diff --git a/3rdparty/astc-encoder/include/astcenc.h b/3rdparty/astc-encoder/include/astcenc.h index dbf4599..c6c8c14 100644 --- a/3rdparty/astc-encoder/include/astcenc.h +++ b/3rdparty/astc-encoder/include/astcenc.h @@ -542,6 +542,14 @@ struct astcenc_config */ float tune_2plane_early_out_limit_correlation; + /** + * @brief The config enable for the mode0 fast-path search. + * + * If this is set to TUNE_MIN_TEXELS_MODE0 or higher then the early-out fast mode0 + * search is enabled. This option is ineffective for 3D block sizes. + */ + float tune_search_mode0_enable; + #if defined(ASTCENC_DIAGNOSTICS) /** * @brief The path to save the diagnostic trace data to. diff --git a/3rdparty/astc-encoder/source/astcenc_color_quantize.cpp b/3rdparty/astc-encoder/source/astcenc_color_quantize.cpp index b0fec7a..df17cac 100644 --- a/3rdparty/astc-encoder/source/astcenc_color_quantize.cpp +++ b/3rdparty/astc-encoder/source/astcenc_color_quantize.cpp @@ -40,6 +40,27 @@ #include "astcenc_internal.h" +/** + * @brief Compute the error of an LDR RGB or RGBA encoding. + * + * @param uquant0 The original endpoint 0 color. + * @param uquant1 The original endpoint 1 color. + * @param quant0 The unpacked quantized endpoint 0 color. + * @param quant1 The unpacked quantized endpoint 1 color. + * + * @return The MSE of the encoding. + */ +static float get_rgba_encoding_error( + vfloat4 uquant0, + vfloat4 uquant1, + vint4 quant0, + vint4 quant1 +) { + vfloat4 error0 = uquant0 - int_to_float(quant0); + vfloat4 error1 = uquant1 - int_to_float(quant1); + return hadd_s(error0 * error0 + error1 * error1); +} + /** * @brief Determine the quantized value given a quantization level. * @@ -56,6 +77,26 @@ static inline uint8_t quant_color( return color_unquant_to_uquant_tables[quant_level - QUANT_6][index]; } +/** + * @brief Determine the quantized value given a quantization level. + * + * @param quant_level The quantization level to use. + * @param value The value to convert. This must be in the 0-255 range. + * + * @return The unpacked quantized value, returned in 0-255 range. + */ +static inline vint4 quant_color3( + quant_method quant_level, + vint4 value +) { + vint4 index = value * 2 + 1; + return vint4( + color_unquant_to_uquant_tables[quant_level - QUANT_6][index.lane<0>()], + color_unquant_to_uquant_tables[quant_level - QUANT_6][index.lane<1>()], + color_unquant_to_uquant_tables[quant_level - QUANT_6][index.lane<2>()], + 0); +} + /** * @brief Determine the quantized value given a quantization level and residual. * @@ -83,6 +124,35 @@ static inline uint8_t quant_color( return color_unquant_to_uquant_tables[quant_level - QUANT_6][index]; } +/** + * @brief Determine the quantized value given a quantization level and residual. + * + * @param quant_level The quantization level to use. + * @param value The value to convert. This must be in the 0-255 range. + * @param valuef The original value before rounding, used to compute a residual. + * + * @return The unpacked quantized value, returned in 0-255 range. + */ +static inline vint4 quant_color3( + quant_method quant_level, + vint4 value, + vfloat4 valuef +) { + vint4 index = value * 2; + + // Compute the residual to determine if we should round down or up ties. + // Test should be residual >= 0, but empirical testing shows small bias helps. + vfloat4 residual = valuef - int_to_float(value); + vmask4 mask = residual >= vfloat4(-0.1f); + index = select(index, index + 1, mask); + + return vint4( + color_unquant_to_uquant_tables[quant_level - QUANT_6][index.lane<0>()], + color_unquant_to_uquant_tables[quant_level - QUANT_6][index.lane<1>()], + color_unquant_to_uquant_tables[quant_level - QUANT_6][index.lane<2>()], + 0); +} + /** * @brief Quantize an LDR RGB color. * @@ -92,47 +162,33 @@ static inline uint8_t quant_color( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r0, r1, g0, g1, b0, b1). + * @param[out] color0_out The output quantized color0 endpoint. + * @param[out] color1_out The output quantized color1 endpoint. * @param quant_level The quantization level to use. */ static void quantize_rgb( vfloat4 color0, vfloat4 color1, - uint8_t output[6], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - float scale = 1.0f / 257.0f; + vint4 color0i, color1i; + vfloat4 nudge(0.2f); - float r0 = astc::clamp255f(color0.lane<0>() * scale); - float g0 = astc::clamp255f(color0.lane<1>() * scale); - float b0 = astc::clamp255f(color0.lane<2>() * scale); - - float r1 = astc::clamp255f(color1.lane<0>() * scale); - float g1 = astc::clamp255f(color1.lane<1>() * scale); - float b1 = astc::clamp255f(color1.lane<2>() * scale); - - int ri0, gi0, bi0, ri1, gi1, bi1; - float rgb0_addon = 0.0f; - float rgb1_addon = 0.0f; do { - ri0 = quant_color(quant_level, astc::max(astc::flt2int_rtn(r0 + rgb0_addon), 0), r0 + rgb0_addon); - gi0 = quant_color(quant_level, astc::max(astc::flt2int_rtn(g0 + rgb0_addon), 0), g0 + rgb0_addon); - bi0 = quant_color(quant_level, astc::max(astc::flt2int_rtn(b0 + rgb0_addon), 0), b0 + rgb0_addon); - ri1 = quant_color(quant_level, astc::min(astc::flt2int_rtn(r1 + rgb1_addon), 255), r1 + rgb1_addon); - gi1 = quant_color(quant_level, astc::min(astc::flt2int_rtn(g1 + rgb1_addon), 255), g1 + rgb1_addon); - bi1 = quant_color(quant_level, astc::min(astc::flt2int_rtn(b1 + rgb1_addon), 255), b1 + rgb1_addon); + vint4 color0q = max(float_to_int_rtn(color0), vint4(0)); + color0i = quant_color3(quant_level, color0q, color0); + color0 = color0 - nudge; - rgb0_addon -= 0.2f; - rgb1_addon += 0.2f; - } while (ri0 + gi0 + bi0 > ri1 + gi1 + bi1); + vint4 color1q = min(float_to_int_rtn(color1), vint4(255)); + color1i = quant_color3(quant_level, color1q, color1); + color1 = color1 + nudge; + } while (hadd_rgb_s(color0i) > hadd_rgb_s(color1i)); - output[0] = static_cast(ri0); - output[1] = static_cast(ri1); - output[2] = static_cast(gi0); - output[3] = static_cast(gi1); - output[4] = static_cast(bi0); - output[5] = static_cast(bi1); + color0_out = color0i; + color1_out = color1i; } /** @@ -145,24 +201,24 @@ static void quantize_rgb( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r0, r1, g0, g1, b0, b1, a0, a1). + * @param[out] color0_out The output quantized color0 endpoint. + * @param[out] color1_out The output quantized color1 endpoint. * @param quant_level The quantization level to use. */ static void quantize_rgba( vfloat4 color0, vfloat4 color1, - uint8_t output[8], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - float scale = 1.0f / 257.0f; + quantize_rgb(color0, color1, color0_out, color1_out, quant_level); - float a0 = astc::clamp255f(color0.lane<3>() * scale); - float a1 = astc::clamp255f(color1.lane<3>() * scale); + float a0 = color0.lane<3>(); + float a1 = color1.lane<3>(); - output[6] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); - output[7] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); - - quantize_rgb(color0, color1, output, quant_level); + color0_out.set_lane<3>(quant_color(quant_level, astc::flt2int_rtn(a0), a0)); + color1_out.set_lane<3>(quant_color(quant_level, astc::flt2int_rtn(a1), a1)); } /** @@ -172,7 +228,8 @@ static void quantize_rgba( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r1, r0, g1, g0, b1, b0). + * @param[out] color0_out The output quantized color0 endpoint. + * @param[out] color1_out The output quantized color1 endpoint. * @param quant_level The quantization level to use. * * @return Returns @c false on failure, @c true on success. @@ -180,54 +237,35 @@ static void quantize_rgba( static bool try_quantize_rgb_blue_contract( vfloat4 color0, vfloat4 color1, - uint8_t output[6], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - float scale = 1.0f / 257.0f; + // Apply inverse blue-contraction + color0 += color0 - color0.swz<2, 2, 2, 3>(); + color1 += color1 - color1.swz<2, 2, 2, 3>(); - float r0 = color0.lane<0>() * scale; - float g0 = color0.lane<1>() * scale; - float b0 = color0.lane<2>() * scale; - - float r1 = color1.lane<0>() * scale; - float g1 = color1.lane<1>() * scale; - float b1 = color1.lane<2>() * scale; - - // Apply inverse blue-contraction. This can produce an overflow; which means BC cannot be used. - r0 += (r0 - b0); - g0 += (g0 - b0); - r1 += (r1 - b1); - g1 += (g1 - b1); - - if (r0 < 0.0f || r0 > 255.0f || g0 < 0.0f || g0 > 255.0f || b0 < 0.0f || b0 > 255.0f || - r1 < 0.0f || r1 > 255.0f || g1 < 0.0f || g1 > 255.0f || b1 < 0.0f || b1 > 255.0f) + // If anything overflows BC cannot be used + vmask4 color0_error = (color0 < vfloat4(0.0f)) | (color0 > vfloat4(255.0f)); + vmask4 color1_error = (color1 < vfloat4(0.0f)) | (color1 > vfloat4(255.0f)); + if (any(color0_error | color1_error)) { return false; } - // Quantize the inverse-blue-contracted color - int ri0 = quant_color(quant_level, astc::flt2int_rtn(r0), r0); - int gi0 = quant_color(quant_level, astc::flt2int_rtn(g0), g0); - int bi0 = quant_color(quant_level, astc::flt2int_rtn(b0), b0); + // Quantize the inverse blue-contracted color + vint4 color0i = quant_color3(quant_level, float_to_int_rtn(color0), color0); + vint4 color1i = quant_color3(quant_level, float_to_int_rtn(color1), color1); - int ri1 = quant_color(quant_level, astc::flt2int_rtn(r1), r1); - int gi1 = quant_color(quant_level, astc::flt2int_rtn(g1), g1); - int bi1 = quant_color(quant_level, astc::flt2int_rtn(b1), b1); - - // If color #1 is not larger than color #0 then blue-contraction cannot be used. Note that - // blue-contraction and quantization change this order, which is why we must test afterwards. - if (ri1 + gi1 + bi1 <= ri0 + gi0 + bi0) + // If color #1 is not larger than color #0 then blue-contraction cannot be used + // We must test afterwards because quantization can change the order + if (hadd_rgb_s(color1i) <= hadd_rgb_s(color0i)) { return false; } - output[0] = static_cast(ri1); - output[1] = static_cast(ri0); - output[2] = static_cast(gi1); - output[3] = static_cast(gi0); - output[4] = static_cast(bi1); - output[5] = static_cast(bi0); - + color0_out = color1i; + color1_out = color0i; return true; } @@ -238,7 +276,8 @@ static bool try_quantize_rgb_blue_contract( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r1, r0, g1, g0, b1, b0, a1, a0). + * @param[out] color0_out The output quantized color0 endpoint. + * @param[out] color1_out The output quantized color1 endpoint. * @param quant_level The quantization level to use. * * @return Returns @c false on failure, @c true on success. @@ -246,18 +285,22 @@ static bool try_quantize_rgb_blue_contract( static bool try_quantize_rgba_blue_contract( vfloat4 color0, vfloat4 color1, - uint8_t output[8], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - float scale = 1.0f / 257.0f; + if (try_quantize_rgb_blue_contract(color0, color1, color0_out, color1_out, quant_level)) + { + float a0 = color0.lane<3>(); + float a1 = color1.lane<3>(); - float a0 = astc::clamp255f(color0.lane<3>() * scale); - float a1 = astc::clamp255f(color1.lane<3>() * scale); + color0_out.set_lane<3>(quant_color(quant_level, astc::flt2int_rtn(a1), a1)); + color1_out.set_lane<3>(quant_color(quant_level, astc::flt2int_rtn(a0), a0)); - output[6] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); - output[7] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); + return true; + } - return try_quantize_rgb_blue_contract(color0, color1, output, quant_level); + return false; } /** @@ -269,7 +312,8 @@ static bool try_quantize_rgba_blue_contract( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r0, r1, g0, g1, b0, b1). + * @param[out] color0_out The output quantized color0 endpoint. + * @param[out] color1_out The output quantized color1 endpoint. * @param quant_level The quantization level to use. * * @return Returns @c false on failure, @c true on success. @@ -277,85 +321,54 @@ static bool try_quantize_rgba_blue_contract( static bool try_quantize_rgb_delta( vfloat4 color0, vfloat4 color1, - uint8_t output[6], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - float scale = 1.0f / 257.0f; - - float r0 = astc::clamp255f(color0.lane<0>() * scale); - float g0 = astc::clamp255f(color0.lane<1>() * scale); - float b0 = astc::clamp255f(color0.lane<2>() * scale); - - float r1 = astc::clamp255f(color1.lane<0>() * scale); - float g1 = astc::clamp255f(color1.lane<1>() * scale); - float b1 = astc::clamp255f(color1.lane<2>() * scale); - - // Transform r0 to unorm9 - int r0a = astc::flt2int_rtn(r0); - int g0a = astc::flt2int_rtn(g0); - int b0a = astc::flt2int_rtn(b0); - - r0a <<= 1; - g0a <<= 1; - b0a <<= 1; + // Transform color0 to unorm9 + vint4 color0a = float_to_int_rtn(color0); + color0.set_lane<3>(0.0f); + color0a = lsl<1>(color0a); // Mask off the top bit - int r0b = r0a & 0xFF; - int g0b = g0a & 0xFF; - int b0b = b0a & 0xFF; + vint4 color0b = color0a & 0xFF; // Quantize then unquantize in order to get a value that we take differences against - int r0be = quant_color(quant_level, r0b); - int g0be = quant_color(quant_level, g0b); - int b0be = quant_color(quant_level, b0b); - - r0b = r0be | (r0a & 0x100); - g0b = g0be | (g0a & 0x100); - b0b = b0be | (b0a & 0x100); + vint4 color0be = quant_color3(quant_level, color0b); + color0b = color0be | (color0a & 0x100); // Get hold of the second value - int r1d = astc::flt2int_rtn(r1); - int g1d = astc::flt2int_rtn(g1); - int b1d = astc::flt2int_rtn(b1); - - r1d <<= 1; - g1d <<= 1; - b1d <<= 1; + vint4 color1d = float_to_int_rtn(color1); + color1d = lsl<1>(color1d); // ... and take differences - r1d -= r0b; - g1d -= g0b; - b1d -= b0b; + color1d = color1d - color0b; + color1d.set_lane<3>(0); // Check if the difference is too large to be encodable - if (r1d > 63 || g1d > 63 || b1d > 63 || r1d < -64 || g1d < -64 || b1d < -64) + if (any((color1d > vint4(63)) | (color1d < vint4(-64)))) { return false; } // Insert top bit of the base into the offset - r1d &= 0x7F; - g1d &= 0x7F; - b1d &= 0x7F; - - r1d |= (r0b & 0x100) >> 1; - g1d |= (g0b & 0x100) >> 1; - b1d |= (b0b & 0x100) >> 1; + color1d = color1d & 0x7F; + color1d = color1d | lsr<1>(color0b & 0x100); // Then quantize and unquantize; if this causes either top two bits to flip, then encoding fails // since we have then corrupted either the top bit of the base or the sign bit of the offset - int r1de = quant_color(quant_level, r1d); - int g1de = quant_color(quant_level, g1d); - int b1de = quant_color(quant_level, b1d); + vint4 color1de = quant_color3(quant_level, color1d); - if (((r1d ^ r1de) | (g1d ^ g1de) | (b1d ^ b1de)) & 0xC0) + vint4 color_flips = (color1d ^ color1de) & 0xC0; + color_flips.set_lane<3>(0); + if (any(color_flips != vint4::zero())) { return false; } // If the sum of offsets triggers blue-contraction then encoding fails - vint4 ep0(r0be, g0be, b0be, 0); - vint4 ep1(r1de, g1de, b1de, 0); + vint4 ep0 = color0be; + vint4 ep1 = color1de; bit_transfer_signed(ep1, ep0); if (hadd_rgb_s(ep1) < 0) { @@ -369,111 +382,90 @@ static bool try_quantize_rgb_delta( return false; } - output[0] = static_cast(r0be); - output[1] = static_cast(r1de); - output[2] = static_cast(g0be); - output[3] = static_cast(g1de); - output[4] = static_cast(b0be); - output[5] = static_cast(b1de); - + color0_out = color0be; + color1_out = color1de; return true; } +/** + * @brief Try to quantize an LDR RGB color using delta encoding and blue-contraction. + * + * Blue-contraction is only usable if encoded color 1 RGB is larger than color 0 RGB. + * + * @param color0 The input unquantized color0 endpoint. + * @param color1 The input unquantized color1 endpoint. + * @param[out] color0_out The output quantized color0 endpoint. + * @param[out] color1_out The output quantized color1 endpoint. + * @param quant_level The quantization level to use. + * + * @return Returns @c false on failure, @c true on success. + */ static bool try_quantize_rgb_delta_blue_contract( vfloat4 color0, vfloat4 color1, - uint8_t output[6], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { // Note: Switch around endpoint colors already at start - float scale = 1.0f / 257.0f; + std::swap(color0, color1); - float r1 = color0.lane<0>() * scale; - float g1 = color0.lane<1>() * scale; - float b1 = color0.lane<2>() * scale; + // Apply inverse blue-contraction + color0 += color0 - color0.swz<2, 2, 2, 3>(); + color1 += color1 - color1.swz<2, 2, 2, 3>(); - float r0 = color1.lane<0>() * scale; - float g0 = color1.lane<1>() * scale; - float b0 = color1.lane<2>() * scale; - - // Apply inverse blue-contraction. This can produce an overflow; which means BC cannot be used. - r0 += (r0 - b0); - g0 += (g0 - b0); - r1 += (r1 - b1); - g1 += (g1 - b1); - - if (r0 < 0.0f || r0 > 255.0f || g0 < 0.0f || g0 > 255.0f || b0 < 0.0f || b0 > 255.0f || - r1 < 0.0f || r1 > 255.0f || g1 < 0.0f || g1 > 255.0f || b1 < 0.0f || b1 > 255.0f) + // If anything overflows BC cannot be used + vmask4 color0_error = (color0 < vfloat4(0.0f)) | (color0 > vfloat4(255.0f)); + vmask4 color1_error = (color1 < vfloat4(0.0f)) | (color1 > vfloat4(255.0f)); + if (any(color0_error | color1_error)) { return false; } - // Transform r0 to unorm9 - int r0a = astc::flt2int_rtn(r0); - int g0a = astc::flt2int_rtn(g0); - int b0a = astc::flt2int_rtn(b0); - r0a <<= 1; - g0a <<= 1; - b0a <<= 1; + // Transform color0 to unorm9 + vint4 color0a = float_to_int_rtn(color0); + color0.set_lane<3>(0.0f); + color0a = lsl<1>(color0a); // Mask off the top bit - int r0b = r0a & 0xFF; - int g0b = g0a & 0xFF; - int b0b = b0a & 0xFF; + vint4 color0b = color0a & 0xFF; - // Quantize, then unquantize in order to get a value that we take differences against. - int r0be = quant_color(quant_level, r0b); - int g0be = quant_color(quant_level, g0b); - int b0be = quant_color(quant_level, b0b); - - r0b = r0be | (r0a & 0x100); - g0b = g0be | (g0a & 0x100); - b0b = b0be | (b0a & 0x100); + // Quantize then unquantize in order to get a value that we take differences against + vint4 color0be = quant_color3(quant_level, color0b); + color0b = color0be | (color0a & 0x100); // Get hold of the second value - int r1d = astc::flt2int_rtn(r1); - int g1d = astc::flt2int_rtn(g1); - int b1d = astc::flt2int_rtn(b1); + vint4 color1d = float_to_int_rtn(color1); + color1d = lsl<1>(color1d); - r1d <<= 1; - g1d <<= 1; - b1d <<= 1; - - // .. and take differences! - r1d -= r0b; - g1d -= g0b; - b1d -= b0b; + // ... and take differences + color1d = color1d - color0b; + color1d.set_lane<3>(0); // Check if the difference is too large to be encodable - if (r1d > 63 || g1d > 63 || b1d > 63 || r1d < -64 || g1d < -64 || b1d < -64) + if (any((color1d > vint4(63)) | (color1d < vint4(-64)))) { return false; } // Insert top bit of the base into the offset - r1d &= 0x7F; - g1d &= 0x7F; - b1d &= 0x7F; + color1d = color1d & 0x7F; + color1d = color1d | lsr<1>(color0b & 0x100); - r1d |= (r0b & 0x100) >> 1; - g1d |= (g0b & 0x100) >> 1; - b1d |= (b0b & 0x100) >> 1; + // Then quantize and unquantize; if this causes either top two bits to flip, then encoding fails + // since we have then corrupted either the top bit of the base or the sign bit of the offset + vint4 color1de = quant_color3(quant_level, color1d); - // Then quantize and unquantize; if this causes any of the top two bits to flip, - // then encoding fails, since we have then corrupted either the top bit of the base - // or the sign bit of the offset. - int r1de = quant_color(quant_level, r1d); - int g1de = quant_color(quant_level, g1d); - int b1de = quant_color(quant_level, b1d); - - if (((r1d ^ r1de) | (g1d ^ g1de) | (b1d ^ b1de)) & 0xC0) + vint4 color_flips = (color1d ^ color1de) & 0xC0; + color_flips.set_lane<3>(0); + if (any(color_flips != vint4::zero())) { return false; } // If the sum of offsets does not trigger blue-contraction then encoding fails - vint4 ep0(r0be, g0be, b0be, 0); - vint4 ep1(r1de, g1de, b1de, 0); + vint4 ep0 = color0be; + vint4 ep1 = color1de; bit_transfer_signed(ep1, ep0); if (hadd_rgb_s(ep1) >= 0) { @@ -487,13 +479,8 @@ static bool try_quantize_rgb_delta_blue_contract( return false; } - output[0] = static_cast(r0be); - output[1] = static_cast(r1de); - output[2] = static_cast(g0be); - output[3] = static_cast(g1de); - output[4] = static_cast(b0be); - output[5] = static_cast(b1de); - + color0_out = color0be; + color1_out = color1de; return true; } @@ -508,7 +495,8 @@ static bool try_quantize_rgb_delta_blue_contract( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (x, x, x, x, x, x, a0, a1). + * @param[out] color0_out The output quantized color0 endpoint; must preserve lane 0/1/2. + * @param[out] color1_out The output quantized color1 endpoint; must preserve lane 0/1/2. * @param quant_level The quantization level to use. * * @return Returns @c false on failure, @c true on success. @@ -516,13 +504,12 @@ static bool try_quantize_rgb_delta_blue_contract( static bool try_quantize_alpha_delta( vfloat4 color0, vfloat4 color1, - uint8_t output[8], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - float scale = 1.0f / 257.0f; - - float a0 = astc::clamp255f(color0.lane<3>() * scale); - float a1 = astc::clamp255f(color1.lane<3>() * scale); + float a0 = color0.lane<3>(); + float a1 = color1.lane<3>(); int a0a = astc::flt2int_rtn(a0); a0a <<= 1; @@ -561,8 +548,8 @@ static bool try_quantize_alpha_delta( return false; } - output[6] = static_cast(a0be); - output[7] = static_cast(a1de); + color0_out.set_lane<3>(a0be); + color1_out.set_lane<3>(a1de); return true; } @@ -589,13 +576,11 @@ static bool try_quantize_luminance_alpha_delta( uint8_t output[4], quant_method quant_level ) { - float scale = 1.0f / 257.0f; + float l0 = hadd_rgb_s(color0) * (1.0f / 3.0f); + float l1 = hadd_rgb_s(color1) * (1.0f / 3.0f); - float l0 = astc::clamp255f(hadd_rgb_s(color0) * ((1.0f / 3.0f) * scale)); - float l1 = astc::clamp255f(hadd_rgb_s(color1) * ((1.0f / 3.0f) * scale)); - - float a0 = astc::clamp255f(color0.lane<3>() * scale); - float a1 = astc::clamp255f(color1.lane<3>() * scale); + float a0 = color0.lane<3>(); + float a1 = color1.lane<3>(); int l0a = astc::flt2int_rtn(l0); int a0a = astc::flt2int_rtn(a0); @@ -693,7 +678,8 @@ static bool try_quantize_luminance_alpha_delta( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r0, r1, b0, b1, g0, g1, a0, a1). + * @param[out] color0_out The output quantized color0 endpoint + * @param[out] color1_out The output quantized color1 endpoint * @param quant_level The quantization level to use. * * @return Returns @c false on failure, @c true on success. @@ -701,14 +687,14 @@ static bool try_quantize_luminance_alpha_delta( static bool try_quantize_rgba_delta( vfloat4 color0, vfloat4 color1, - uint8_t output[8], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { - return try_quantize_rgb_delta(color0, color1, output, quant_level) && - try_quantize_alpha_delta(color0, color1, output, quant_level); + return try_quantize_rgb_delta(color0, color1, color0_out, color1_out, quant_level) && + try_quantize_alpha_delta(color0, color1, color0_out, color1_out, quant_level); } - /** * @brief Try to quantize an LDR RGBA color using delta and blue contract encoding. * @@ -720,7 +706,8 @@ static bool try_quantize_rgba_delta( * * @param color0 The input unquantized color0 endpoint. * @param color1 The input unquantized color1 endpoint. - * @param[out] output The output endpoints, returned as (r0, r1, b0, b1, g0, g1, a0, a1). + * @param[out] color0_out The output quantized color0 endpoint + * @param[out] color1_out The output quantized color1 endpoint * @param quant_level The quantization level to use. * * @return Returns @c false on failure, @c true on success. @@ -728,12 +715,13 @@ static bool try_quantize_rgba_delta( static bool try_quantize_rgba_delta_blue_contract( vfloat4 color0, vfloat4 color1, - uint8_t output[8], + vint4& color0_out, + vint4& color1_out, quant_method quant_level ) { // Note that we swap the color0 and color1 ordering for alpha to match RGB blue-contract - return try_quantize_rgb_delta_blue_contract(color0, color1, output, quant_level) && - try_quantize_alpha_delta(color1, color0, output, quant_level); + return try_quantize_rgb_delta_blue_contract(color0, color1, color0_out, color1_out, quant_level) && + try_quantize_alpha_delta(color1, color0, color0_out, color1_out, quant_level); } /** @@ -774,6 +762,8 @@ static void quantize_rgbs( /** * @brief Quantize an LDR RGBA color using scale encoding. * + * @param color0 The input unquantized color0 alpha endpoint. + * @param color1 The input unquantized color1 alpha endpoint. * @param color The input unquantized color endpoint and scale factor. * @param[out] output The output endpoints, returned as (r0, g0, b0, s, a0, a1). * @param quant_level The quantization level to use. @@ -785,10 +775,8 @@ static void quantize_rgbs_alpha( uint8_t output[6], quant_method quant_level ) { - float scale = 1.0f / 257.0f; - - float a0 = astc::clamp255f(color0.lane<3>() * scale); - float a1 = astc::clamp255f(color1.lane<3>() * scale); + float a0 = color0.lane<3>(); + float a1 = color1.lane<3>(); output[4] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); output[5] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); @@ -810,13 +798,8 @@ static void quantize_luminance( uint8_t output[2], quant_method quant_level ) { - float scale = 1.0f / 257.0f; - - color0 = color0 * scale; - color1 = color1 * scale; - - float lum0 = astc::clamp255f(hadd_rgb_s(color0) * (1.0f / 3.0f)); - float lum1 = astc::clamp255f(hadd_rgb_s(color1) * (1.0f / 3.0f)); + float lum0 = hadd_rgb_s(color0) * (1.0f / 3.0f); + float lum1 = hadd_rgb_s(color1) * (1.0f / 3.0f); if (lum0 > lum1) { @@ -843,16 +826,11 @@ static void quantize_luminance_alpha( uint8_t output[4], quant_method quant_level ) { - float scale = 1.0f / 257.0f; + float lum0 = hadd_rgb_s(color0) * (1.0f / 3.0f); + float lum1 = hadd_rgb_s(color1) * (1.0f / 3.0f); - color0 = color0 * scale; - color1 = color1 * scale; - - float lum0 = astc::clamp255f(hadd_rgb_s(color0) * (1.0f / 3.0f)); - float lum1 = astc::clamp255f(hadd_rgb_s(color1) * (1.0f / 3.0f)); - - float a0 = astc::clamp255f(color0.lane<3>()); - float a1 = astc::clamp255f(color1.lane<3>()); + float a0 = color0.lane<3>(); + float a1 = color1.lane<3>(); output[0] = quant_color(quant_level, astc::flt2int_rtn(lum0), lum0); output[1] = quant_color(quant_level, astc::flt2int_rtn(lum1), lum1); @@ -1939,58 +1917,170 @@ uint8_t pack_color_endpoints( ) { assert(QUANT_6 <= quant_level && quant_level <= QUANT_256); - // We do not support negative colors - color0 = max(color0, 0.0f); - color1 = max(color1, 0.0f); + // Clamp colors to a valid LDR range + // Note that HDR has a lower max, handled in the conversion functions + color0 = clamp(0.0f, 65535.0f, color0); + color1 = clamp(0.0f, 65535.0f, color1); + + // Pre-scale the LDR value we need to the 0-255 quantizable range + vfloat4 color0_ldr = color0 * (1.0f / 257.0f); + vfloat4 color1_ldr = color1 * (1.0f / 257.0f); uint8_t retval = 0; + float best_error = ERROR_CALC_DEFAULT; + vint4 color0_out, color1_out; + vint4 color0_out2, color1_out2; switch (format) { case FMT_RGB: if (quant_level <= QUANT_160) { - if (try_quantize_rgb_delta_blue_contract(color0, color1, output, quant_level)) + if (try_quantize_rgb_delta_blue_contract(color0_ldr, color1_ldr, color0_out, color1_out, quant_level)) { + vint4 color0_unpack; + vint4 color1_unpack; + rgba_delta_unpack(color0_out, color1_out, color0_unpack, color1_unpack); + retval = FMT_RGB_DELTA; - break; + best_error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); } - if (try_quantize_rgb_delta(color0, color1, output, quant_level)) + + if (try_quantize_rgb_delta(color0_ldr, color1_ldr, color0_out2, color1_out2, quant_level)) { - retval = FMT_RGB_DELTA; - break; + vint4 color0_unpack; + vint4 color1_unpack; + rgba_delta_unpack(color0_out2, color1_out2, color0_unpack, color1_unpack); + + float error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); + if (error < best_error) + { + retval = FMT_RGB_DELTA; + best_error = error; + color0_out = color0_out2; + color1_out = color1_out2; + } } } - if (quant_level < QUANT_256 && try_quantize_rgb_blue_contract(color0, color1, output, quant_level)) + + if (quant_level < QUANT_256) { - retval = FMT_RGB; - break; + if (try_quantize_rgb_blue_contract(color0_ldr, color1_ldr, color0_out2, color1_out2, quant_level)) + { + vint4 color0_unpack; + vint4 color1_unpack; + rgba_unpack(color0_out2, color1_out2, color0_unpack, color1_unpack); + + float error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); + if (error < best_error) + { + retval = FMT_RGB; + best_error = error; + color0_out = color0_out2; + color1_out = color1_out2; + } + } } - quantize_rgb(color0, color1, output, quant_level); - retval = FMT_RGB; + + { + quantize_rgb(color0_ldr, color1_ldr, color0_out2, color1_out2, quant_level); + + vint4 color0_unpack; + vint4 color1_unpack; + rgba_unpack(color0_out2, color1_out2, color0_unpack, color1_unpack); + + float error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); + if (error < best_error) + { + retval = FMT_RGB; + color0_out = color0_out2; + color1_out = color1_out2; + } + } + + // TODO: Can we vectorize this? + output[0] = static_cast(color0_out.lane<0>()); + output[1] = static_cast(color1_out.lane<0>()); + output[2] = static_cast(color0_out.lane<1>()); + output[3] = static_cast(color1_out.lane<1>()); + output[4] = static_cast(color0_out.lane<2>()); + output[5] = static_cast(color1_out.lane<2>()); break; case FMT_RGBA: if (quant_level <= QUANT_160) { - if (try_quantize_rgba_delta_blue_contract(color0, color1, output, quant_level)) + if (try_quantize_rgba_delta_blue_contract(color0_ldr, color1_ldr, color0_out, color1_out, quant_level)) { + vint4 color0_unpack; + vint4 color1_unpack; + rgba_delta_unpack(color0_out, color1_out, color0_unpack, color1_unpack); + retval = FMT_RGBA_DELTA; - break; + best_error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); } - if (try_quantize_rgba_delta(color0, color1, output, quant_level)) + + if (try_quantize_rgba_delta(color0_ldr, color1_ldr, color0_out2, color1_out2, quant_level)) { - retval = FMT_RGBA_DELTA; - break; + vint4 color0_unpack; + vint4 color1_unpack; + rgba_delta_unpack(color0_out2, color1_out2, color0_unpack, color1_unpack); + + float error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); + if (error < best_error) + { + retval = FMT_RGBA_DELTA; + best_error = error; + color0_out = color0_out2; + color1_out = color1_out2; + } } } - if (quant_level < QUANT_256 && try_quantize_rgba_blue_contract(color0, color1, output, quant_level)) + + if (quant_level < QUANT_256) { - retval = FMT_RGBA; - break; + if (try_quantize_rgba_blue_contract(color0_ldr, color1_ldr, color0_out2, color1_out2, quant_level)) + { + vint4 color0_unpack; + vint4 color1_unpack; + rgba_unpack(color0_out2, color1_out2, color0_unpack, color1_unpack); + + float error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); + if (error < best_error) + { + retval = FMT_RGBA; + best_error = error; + color0_out = color0_out2; + color1_out = color1_out2; + } + } } - quantize_rgba(color0, color1, output, quant_level); - retval = FMT_RGBA; + + { + quantize_rgba(color0_ldr, color1_ldr, color0_out2, color1_out2, quant_level); + + vint4 color0_unpack; + vint4 color1_unpack; + rgba_unpack(color0_out2, color1_out2, color0_unpack, color1_unpack); + + float error = get_rgba_encoding_error(color0_ldr, color1_ldr, color0_unpack, color1_unpack); + if (error < best_error) + { + retval = FMT_RGBA; + color0_out = color0_out2; + color1_out = color1_out2; + } + } + + // TODO: Can we vectorize this? + output[0] = static_cast(color0_out.lane<0>()); + output[1] = static_cast(color1_out.lane<0>()); + output[2] = static_cast(color0_out.lane<1>()); + output[3] = static_cast(color1_out.lane<1>()); + output[4] = static_cast(color0_out.lane<2>()); + output[5] = static_cast(color1_out.lane<2>()); + output[6] = static_cast(color0_out.lane<3>()); + output[7] = static_cast(color1_out.lane<3>()); break; case FMT_RGB_SCALE: @@ -2009,7 +2099,7 @@ uint8_t pack_color_endpoints( break; case FMT_RGB_SCALE_ALPHA: - quantize_rgbs_alpha(color0, color1, rgbs_color, output, quant_level); + quantize_rgbs_alpha(color0_ldr, color1_ldr, rgbs_color, output, quant_level); retval = FMT_RGB_SCALE_ALPHA; break; @@ -2025,20 +2115,20 @@ uint8_t pack_color_endpoints( break; case FMT_LUMINANCE: - quantize_luminance(color0, color1, output, quant_level); + quantize_luminance(color0_ldr, color1_ldr, output, quant_level); retval = FMT_LUMINANCE; break; case FMT_LUMINANCE_ALPHA: if (quant_level <= 18) { - if (try_quantize_luminance_alpha_delta(color0, color1, output, quant_level)) + if (try_quantize_luminance_alpha_delta(color0_ldr, color1_ldr, output, quant_level)) { retval = FMT_LUMINANCE_ALPHA_DELTA; break; } } - quantize_luminance_alpha(color0, color1, output, quant_level); + quantize_luminance_alpha(color0_ldr, color1_ldr, output, quant_level); retval = FMT_LUMINANCE_ALPHA; break; diff --git a/3rdparty/astc-encoder/source/astcenc_color_unquantize.cpp b/3rdparty/astc-encoder/source/astcenc_color_unquantize.cpp index d31895a..10fb6bc 100644 --- a/3rdparty/astc-encoder/source/astcenc_color_unquantize.cpp +++ b/3rdparty/astc-encoder/source/astcenc_color_unquantize.cpp @@ -40,15 +40,7 @@ static ASTCENC_SIMD_INLINE vint4 uncontract_color( return select(input, bc0, mask); } -/** - * @brief Unpack an LDR RGBA color that uses delta encoding. - * - * @param input0 The packed endpoint 0 color. - * @param input1 The packed endpoint 1 color deltas. - * @param[out] output0 The unpacked endpoint 0 color. - * @param[out] output1 The unpacked endpoint 1 color. - */ -static void rgba_delta_unpack( +void rgba_delta_unpack( vint4 input0, vint4 input1, vint4& output0, @@ -92,15 +84,7 @@ static void rgb_delta_unpack( output1.set_lane<3>(255); } -/** - * @brief Unpack an LDR RGBA color that uses direct encoding. - * - * @param input0 The packed endpoint 0 color. - * @param input1 The packed endpoint 1 color. - * @param[out] output0 The unpacked endpoint 0 color. - * @param[out] output1 The unpacked endpoint 1 color. - */ -static void rgba_unpack( +void rgba_unpack( vint4 input0, vint4 input1, vint4& output0, diff --git a/3rdparty/astc-encoder/source/astcenc_compress_symbolic.cpp b/3rdparty/astc-encoder/source/astcenc_compress_symbolic.cpp index 41a8558..0c90540 100644 --- a/3rdparty/astc-encoder/source/astcenc_compress_symbolic.cpp +++ b/3rdparty/astc-encoder/source/astcenc_compress_symbolic.cpp @@ -1163,7 +1163,7 @@ static float prepare_block_statistics( void compress_block( const astcenc_contexti& ctx, const image_block& blk, - physical_compressed_block& pcb, + uint8_t pcb[16], compression_working_buffers& tmpbuf) { astcenc_profile decode_mode = ctx.config.profile; @@ -1282,9 +1282,10 @@ void compress_block( static const float errorval_overshoot = 1.0f / ctx.config.tune_mse_overshoot; - // Only enable MODE0 fast path (trial 0) if 2D, and more than 25 texels + // Only enable MODE0 fast path if enabled + // Never enable for 3D blocks as no "always" block modes are available int start_trial = 1; - if ((bsd.texel_count >= TUNE_MIN_TEXELS_MODE0_FASTPATH) && (bsd.zdim == 1)) + if ((ctx.config.tune_search_mode0_enable >= TUNE_MIN_SEARCH_MODE0) && (bsd.zdim == 1)) { start_trial = 0; } diff --git a/3rdparty/astc-encoder/source/astcenc_decompress_symbolic.cpp b/3rdparty/astc-encoder/source/astcenc_decompress_symbolic.cpp index 39e5525..dd331a9 100644 --- a/3rdparty/astc-encoder/source/astcenc_decompress_symbolic.cpp +++ b/3rdparty/astc-encoder/source/astcenc_decompress_symbolic.cpp @@ -104,10 +104,10 @@ void unpack_weights( if (!is_dual_plane) { // Build full 64-entry weight lookup table - vint4 tab0(reinterpret_cast(scb.weights + 0)); - vint4 tab1(reinterpret_cast(scb.weights + 16)); - vint4 tab2(reinterpret_cast(scb.weights + 32)); - vint4 tab3(reinterpret_cast(scb.weights + 48)); + vint4 tab0 = vint4::load(scb.weights + 0); + vint4 tab1 = vint4::load(scb.weights + 16); + vint4 tab2 = vint4::load(scb.weights + 32); + vint4 tab3 = vint4::load(scb.weights + 48); vint tab0p, tab1p, tab2p, tab3p; vtable_prepare(tab0, tab1, tab2, tab3, tab0p, tab1p, tab2p, tab3p); @@ -134,14 +134,14 @@ void unpack_weights( { // Build a 32-entry weight lookup table per plane // Plane 1 - vint4 tab0_plane1(reinterpret_cast(scb.weights + 0)); - vint4 tab1_plane1(reinterpret_cast(scb.weights + 16)); + vint4 tab0_plane1 = vint4::load(scb.weights + 0); + vint4 tab1_plane1 = vint4::load(scb.weights + 16); vint tab0_plane1p, tab1_plane1p; vtable_prepare(tab0_plane1, tab1_plane1, tab0_plane1p, tab1_plane1p); // Plane 2 - vint4 tab0_plane2(reinterpret_cast(scb.weights + 32)); - vint4 tab1_plane2(reinterpret_cast(scb.weights + 48)); + vint4 tab0_plane2 = vint4::load(scb.weights + 32); + vint4 tab1_plane2 = vint4::load(scb.weights + 48); vint tab0_plane2p, tab1_plane2p; vtable_prepare(tab0_plane2, tab1_plane2, tab0_plane2p, tab1_plane2p); diff --git a/3rdparty/astc-encoder/source/astcenc_entry.cpp b/3rdparty/astc-encoder/source/astcenc_entry.cpp index e53762c..03cf6a8 100644 --- a/3rdparty/astc-encoder/source/astcenc_entry.cpp +++ b/3rdparty/astc-encoder/source/astcenc_entry.cpp @@ -55,6 +55,7 @@ struct astcenc_preset_config float tune_2partition_early_out_limit_factor; float tune_3partition_early_out_limit_factor; float tune_2plane_early_out_limit_correlation; + float tune_search_mode0_enable; }; /** @@ -63,22 +64,22 @@ struct astcenc_preset_config static const std::array preset_configs_high {{ { ASTCENC_PRE_FASTEST, - 2, 10, 6, 4, 43, 2, 2, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.85f + 2, 10, 6, 4, 43, 2, 2, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.85f, 0.0f }, { ASTCENC_PRE_FAST, - 3, 18, 10, 8, 55, 3, 3, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.90f + 3, 18, 10, 8, 55, 3, 3, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.90f, 0.0f }, { ASTCENC_PRE_MEDIUM, - 4, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 2.5f, 1.1f, 1.05f, 0.95f + 4, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 2.5f, 1.1f, 1.05f, 0.95f, 0.0f }, { ASTCENC_PRE_THOROUGH, - 4, 82, 60, 30, 94, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.35f, 1.15f, 0.97f + 4, 82, 60, 30, 94, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.35f, 1.15f, 0.97f, 0.0f }, { ASTCENC_PRE_VERYTHOROUGH, - 4, 256, 128, 64, 98, 4, 6, 20, 14, 8, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f + 4, 256, 128, 64, 98, 4, 6, 8, 6, 4, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f, 0.0f }, { ASTCENC_PRE_EXHAUSTIVE, - 4, 512, 512, 512, 100, 4, 8, 32, 32, 32, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f + 4, 512, 512, 512, 100, 4, 8, 8, 8, 8, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f, 0.0f } }}; @@ -88,22 +89,22 @@ static const std::array preset_configs_high {{ static const std::array preset_configs_mid {{ { ASTCENC_PRE_FASTEST, - 2, 10, 6, 4, 43, 2, 2, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.80f + 2, 10, 6, 4, 43, 2, 2, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.80f, 1.0f }, { ASTCENC_PRE_FAST, - 3, 18, 12, 10, 55, 3, 3, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.85f + 3, 18, 12, 10, 55, 3, 3, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.85f, 1.0f }, { ASTCENC_PRE_MEDIUM, - 4, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 3.0f, 1.1f, 1.05f, 0.90f + 3, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 3.0f, 1.1f, 1.05f, 0.90f, 1.0f }, { ASTCENC_PRE_THOROUGH, - 4, 82, 60, 30, 94, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.4f, 1.2f, 0.95f + 4, 82, 60, 30, 94, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.4f, 1.2f, 0.95f, 0.0f }, { ASTCENC_PRE_VERYTHOROUGH, - 4, 256, 128, 64, 98, 4, 6, 12, 8, 3, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f + 4, 256, 128, 64, 98, 4, 6, 8, 6, 3, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f, 0.0f }, { ASTCENC_PRE_EXHAUSTIVE, - 4, 256, 256, 256, 100, 4, 8, 32, 32, 32, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f + 4, 256, 256, 256, 100, 4, 8, 8, 8, 8, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f, 0.0f } }}; @@ -113,22 +114,22 @@ static const std::array preset_configs_mid {{ static const std::array preset_configs_low {{ { ASTCENC_PRE_FASTEST, - 2, 10, 6, 4, 40, 2, 2, 2, 2, 2, 85.0f, 63.0f, 3.5f, 1.0f, 1.0f, 0.80f + 2, 10, 6, 4, 40, 2, 2, 2, 2, 2, 85.0f, 63.0f, 3.5f, 1.0f, 1.0f, 0.80f, 1.0f }, { ASTCENC_PRE_FAST, - 2, 18, 12, 10, 55, 3, 3, 2, 2, 2, 85.0f, 63.0f, 3.5f, 1.0f, 1.0f, 0.85f + 2, 18, 12, 10, 55, 3, 3, 2, 2, 2, 85.0f, 63.0f, 3.5f, 1.0f, 1.0f, 0.85f, 1.0f }, { ASTCENC_PRE_MEDIUM, - 3, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 3.5f, 1.1f, 1.05f, 0.90f + 3, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 3.5f, 1.1f, 1.05f, 0.90f, 1.0f }, { ASTCENC_PRE_THOROUGH, - 4, 82, 60, 30, 93, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.3f, 1.2f, 0.97f + 4, 82, 60, 30, 93, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.3f, 1.2f, 0.97f, 1.0f }, { ASTCENC_PRE_VERYTHOROUGH, - 4, 256, 128, 64, 98, 4, 6, 9, 5, 2, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f + 4, 256, 128, 64, 98, 4, 6, 8, 5, 2, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f, 1.0f }, { ASTCENC_PRE_EXHAUSTIVE, - 4, 256, 256, 256, 100, 4, 8, 32, 32, 32, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f + 4, 256, 256, 256, 100, 4, 8, 8, 8, 8, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f, 1.0f } }}; @@ -504,10 +505,10 @@ astcenc_error astcenc_config_init( config.tune_4partition_index_limit = (*preset_configs)[start].tune_4partition_index_limit; config.tune_block_mode_limit = (*preset_configs)[start].tune_block_mode_limit; config.tune_refinement_limit = (*preset_configs)[start].tune_refinement_limit; - config.tune_candidate_limit = astc::min((*preset_configs)[start].tune_candidate_limit, TUNE_MAX_TRIAL_CANDIDATES); - config.tune_2partitioning_candidate_limit = astc::min((*preset_configs)[start].tune_2partitioning_candidate_limit, TUNE_MAX_PARTITIONING_CANDIDATES); - config.tune_3partitioning_candidate_limit = astc::min((*preset_configs)[start].tune_3partitioning_candidate_limit, TUNE_MAX_PARTITIONING_CANDIDATES); - config.tune_4partitioning_candidate_limit = astc::min((*preset_configs)[start].tune_4partitioning_candidate_limit, TUNE_MAX_PARTITIONING_CANDIDATES); + config.tune_candidate_limit = (*preset_configs)[start].tune_candidate_limit; + config.tune_2partitioning_candidate_limit = (*preset_configs)[start].tune_2partitioning_candidate_limit; + config.tune_3partitioning_candidate_limit = (*preset_configs)[start].tune_3partitioning_candidate_limit; + config.tune_4partitioning_candidate_limit = (*preset_configs)[start].tune_4partitioning_candidate_limit; config.tune_db_limit = astc::max((*preset_configs)[start].tune_db_limit_a_base - 35 * ltexels, (*preset_configs)[start].tune_db_limit_b_base - 19 * ltexels); @@ -516,6 +517,7 @@ astcenc_error astcenc_config_init( config.tune_2partition_early_out_limit_factor = (*preset_configs)[start].tune_2partition_early_out_limit_factor; config.tune_3partition_early_out_limit_factor = (*preset_configs)[start].tune_3partition_early_out_limit_factor; config.tune_2plane_early_out_limit_correlation = (*preset_configs)[start].tune_2plane_early_out_limit_correlation; + config.tune_search_mode0_enable = (*preset_configs)[start].tune_search_mode0_enable; } // Start and end node are not the same - so interpolate between them else @@ -542,14 +544,10 @@ astcenc_error astcenc_config_init( config.tune_4partition_index_limit = LERPI(tune_4partition_index_limit); config.tune_block_mode_limit = LERPI(tune_block_mode_limit); config.tune_refinement_limit = LERPI(tune_refinement_limit); - config.tune_candidate_limit = astc::min(LERPUI(tune_candidate_limit), - TUNE_MAX_TRIAL_CANDIDATES); - config.tune_2partitioning_candidate_limit = astc::min(LERPUI(tune_2partitioning_candidate_limit), - BLOCK_MAX_PARTITIONINGS); - config.tune_3partitioning_candidate_limit = astc::min(LERPUI(tune_3partitioning_candidate_limit), - BLOCK_MAX_PARTITIONINGS); - config.tune_4partitioning_candidate_limit = astc::min(LERPUI(tune_4partitioning_candidate_limit), - BLOCK_MAX_PARTITIONINGS); + config.tune_candidate_limit = LERPUI(tune_candidate_limit); + config.tune_2partitioning_candidate_limit = LERPUI(tune_2partitioning_candidate_limit); + config.tune_3partitioning_candidate_limit = LERPUI(tune_3partitioning_candidate_limit); + config.tune_4partitioning_candidate_limit = LERPUI(tune_4partitioning_candidate_limit); config.tune_db_limit = astc::max(LERP(tune_db_limit_a_base) - 35 * ltexels, LERP(tune_db_limit_b_base) - 19 * ltexels); @@ -558,6 +556,7 @@ astcenc_error astcenc_config_init( config.tune_2partition_early_out_limit_factor = LERP(tune_2partition_early_out_limit_factor); config.tune_3partition_early_out_limit_factor = LERP(tune_3partition_early_out_limit_factor); config.tune_2plane_early_out_limit_correlation = LERP(tune_2plane_early_out_limit_correlation); + config.tune_search_mode0_enable = LERP(tune_search_mode0_enable); #undef LERP #undef LERPI #undef LERPUI @@ -585,6 +584,7 @@ astcenc_error astcenc_config_init( case ASTCENC_PRF_HDR_RGB_LDR_A: case ASTCENC_PRF_HDR: config.tune_db_limit = 999.0f; + config.tune_search_mode0_enable = 0.0f; break; default: return ASTCENC_ERR_BAD_PROFILE; @@ -914,8 +914,7 @@ static void compress_image( int offset = ((z * yblocks + y) * xblocks + x) * 16; uint8_t *bp = buffer + offset; - physical_compressed_block* pcb = reinterpret_cast(bp); - compress_block(ctx, blk, *pcb, temp_buffers); + compress_block(ctx, blk, bp, temp_buffers); } ctxo.manage_compress.complete_task_assignment(count); @@ -1182,10 +1181,9 @@ astcenc_error astcenc_decompress_image( unsigned int offset = (((z * yblocks + y) * xblocks) + x) * 16; const uint8_t* bp = data + offset; - const physical_compressed_block& pcb = *reinterpret_cast(bp); symbolic_compressed_block scb; - physical_to_symbolic(*ctx->bsd, pcb, scb); + physical_to_symbolic(*ctx->bsd, bp, scb); decompress_symbolic_block(ctx->config.profile, *ctx->bsd, x * block_x, y * block_y, z * block_z, @@ -1224,9 +1222,8 @@ astcenc_error astcenc_get_block_info( astcenc_contexti* ctx = &ctxo->context; // Decode the compressed data into a symbolic form - const physical_compressed_block&pcb = *reinterpret_cast(data); symbolic_compressed_block scb; - physical_to_symbolic(*ctx->bsd, pcb, scb); + physical_to_symbolic(*ctx->bsd, data, scb); // Fetch the appropriate partition and decimation tables block_size_descriptor& bsd = *ctx->bsd; diff --git a/3rdparty/astc-encoder/source/astcenc_find_best_partitioning.cpp b/3rdparty/astc-encoder/source/astcenc_find_best_partitioning.cpp index 789c396..bfbcc35 100644 --- a/3rdparty/astc-encoder/source/astcenc_find_best_partitioning.cpp +++ b/3rdparty/astc-encoder/source/astcenc_find_best_partitioning.cpp @@ -250,13 +250,16 @@ static void kmeans_update( * * @return The number of bit mismatches. */ -static inline unsigned int partition_mismatch2( +static inline uint8_t partition_mismatch2( const uint64_t a[2], const uint64_t b[2] ) { int v1 = popcount(a[0] ^ b[0]) + popcount(a[1] ^ b[1]); int v2 = popcount(a[0] ^ b[1]) + popcount(a[1] ^ b[0]); - return astc::min(v1, v2); + + // Divide by 2 because XOR always counts errors twice, once when missing + // in the expected position, and again when present in the wrong partition + return static_cast(astc::min(v1, v2) / 2); } /** @@ -267,7 +270,7 @@ static inline unsigned int partition_mismatch2( * * @return The number of bit mismatches. */ -static inline unsigned int partition_mismatch3( +static inline uint8_t partition_mismatch3( const uint64_t a[3], const uint64_t b[3] ) { @@ -295,7 +298,9 @@ static inline unsigned int partition_mismatch3( int s5 = p11 + p20; int v2 = astc::min(s4, s5) + p02; - return astc::min(v0, v1, v2); + // Divide by 2 because XOR always counts errors twice, once when missing + // in the expected position, and again when present in the wrong partition + return static_cast(astc::min(v0, v1, v2) / 2); } /** @@ -306,7 +311,7 @@ static inline unsigned int partition_mismatch3( * * @return The number of bit mismatches. */ -static inline unsigned int partition_mismatch4( +static inline uint8_t partition_mismatch4( const uint64_t a[4], const uint64_t b[4] ) { @@ -342,7 +347,9 @@ static inline unsigned int partition_mismatch4( int v2 = p02 + astc::min(p11 + mx03, p10 + mx13, p13 + mx01); int v3 = p03 + astc::min(p11 + mx02, p12 + mx01, p10 + mx12); - return astc::min(v0, v1, v2, v3); + // Divide by 2 because XOR always counts errors twice, once when missing + // in the expected position, and again when present in the wrong partition + return static_cast(astc::min(v0, v1, v2, v3) / 2); } using mismatch_dispatch = unsigned int (*)(const uint64_t*, const uint64_t*); @@ -359,7 +366,7 @@ static void count_partition_mismatch_bits( const block_size_descriptor& bsd, unsigned int partition_count, const uint64_t bitmaps[BLOCK_MAX_PARTITIONS], - unsigned int mismatch_counts[BLOCK_MAX_PARTITIONINGS] + uint8_t mismatch_counts[BLOCK_MAX_PARTITIONINGS] ) { unsigned int active_count = bsd.partitioning_count_selected[partition_count - 1]; promise(active_count > 0); @@ -369,6 +376,8 @@ static void count_partition_mismatch_bits( for (unsigned int i = 0; i < active_count; i++) { mismatch_counts[i] = partition_mismatch2(bitmaps, bsd.coverage_bitmaps_2[i]); + assert(mismatch_counts[i] < BLOCK_MAX_KMEANS_TEXELS); + assert(mismatch_counts[i] < bsd.texel_count); } } else if (partition_count == 3) @@ -376,6 +385,8 @@ static void count_partition_mismatch_bits( for (unsigned int i = 0; i < active_count; i++) { mismatch_counts[i] = partition_mismatch3(bitmaps, bsd.coverage_bitmaps_3[i]); + assert(mismatch_counts[i] < BLOCK_MAX_KMEANS_TEXELS); + assert(mismatch_counts[i] < bsd.texel_count); } } else @@ -383,6 +394,8 @@ static void count_partition_mismatch_bits( for (unsigned int i = 0; i < active_count; i++) { mismatch_counts[i] = partition_mismatch4(bitmaps, bsd.coverage_bitmaps_4[i]); + assert(mismatch_counts[i] < BLOCK_MAX_KMEANS_TEXELS); + assert(mismatch_counts[i] < bsd.texel_count); } } } @@ -397,12 +410,13 @@ static void count_partition_mismatch_bits( * @return The number of active partitions in this selection. */ static unsigned int get_partition_ordering_by_mismatch_bits( + unsigned int texel_count, unsigned int partitioning_count, - const unsigned int mismatch_count[BLOCK_MAX_PARTITIONINGS], - unsigned int partition_ordering[BLOCK_MAX_PARTITIONINGS] + const uint8_t mismatch_count[BLOCK_MAX_PARTITIONINGS], + uint16_t partition_ordering[BLOCK_MAX_PARTITIONINGS] ) { promise(partitioning_count > 0); - unsigned int mscount[256] { 0 }; + uint16_t mscount[BLOCK_MAX_KMEANS_TEXELS] { 0 }; // Create the histogram of mismatch counts for (unsigned int i = 0; i < partitioning_count; i++) @@ -410,16 +424,14 @@ static unsigned int get_partition_ordering_by_mismatch_bits( mscount[mismatch_count[i]]++; } - unsigned int active_count = partitioning_count - mscount[255]; - // Create a running sum from the histogram array // Cells store previous values only; i.e. exclude self after sum - unsigned int summa = 0; - for (unsigned int i = 0; i < 256; i++) + unsigned int sum = 0; + for (unsigned int i = 0; i < texel_count; i++) { - unsigned int cnt = mscount[i]; - mscount[i] = summa; - summa += cnt; + uint16_t cnt = mscount[i]; + mscount[i] = sum; + sum += cnt; } // Use the running sum as the index, incrementing after read to allow @@ -427,10 +439,10 @@ static unsigned int get_partition_ordering_by_mismatch_bits( for (unsigned int i = 0; i < partitioning_count; i++) { unsigned int idx = mscount[mismatch_count[i]]++; - partition_ordering[idx] = i; + partition_ordering[idx] = static_cast(i); } - return active_count; + return partitioning_count; } /** @@ -447,7 +459,7 @@ static unsigned int compute_kmeans_partition_ordering( const block_size_descriptor& bsd, const image_block& blk, unsigned int partition_count, - unsigned int partition_ordering[BLOCK_MAX_PARTITIONINGS] + uint16_t partition_ordering[BLOCK_MAX_PARTITIONINGS] ) { vfloat4 cluster_centers[BLOCK_MAX_PARTITIONS]; uint8_t texel_partitions[BLOCK_MAX_TEXELS]; @@ -478,11 +490,12 @@ static unsigned int compute_kmeans_partition_ordering( } // Count the mismatch between the block and the format's partition tables - unsigned int mismatch_counts[BLOCK_MAX_PARTITIONINGS]; + uint8_t mismatch_counts[BLOCK_MAX_PARTITIONINGS]; count_partition_mismatch_bits(bsd, partition_count, bitmaps, mismatch_counts); // Sort the partitions based on the number of mismatched bits return get_partition_ordering_by_mismatch_bits( + texels_to_process, bsd.partitioning_count_selected[partition_count - 1], mismatch_counts, partition_ordering); } @@ -565,7 +578,7 @@ unsigned int find_best_partition_candidates( weight_imprecision_estim = weight_imprecision_estim * weight_imprecision_estim; - unsigned int partition_sequence[BLOCK_MAX_PARTITIONINGS]; + uint16_t partition_sequence[BLOCK_MAX_PARTITIONINGS]; unsigned int sequence_len = compute_kmeans_partition_ordering(bsd, blk, partition_count, partition_sequence); partition_search_limit = astc::min(partition_search_limit, sequence_len); requested_candidates = astc::min(partition_search_limit, requested_candidates); diff --git a/3rdparty/astc-encoder/source/astcenc_ideal_endpoints_and_weights.cpp b/3rdparty/astc-encoder/source/astcenc_ideal_endpoints_and_weights.cpp index 5145e08..89ec9dc 100644 --- a/3rdparty/astc-encoder/source/astcenc_ideal_endpoints_and_weights.cpp +++ b/3rdparty/astc-encoder/source/astcenc_ideal_endpoints_and_weights.cpp @@ -1023,7 +1023,7 @@ void compute_quantized_weights_for_decimation( // safe data in compute_ideal_weights_for_decimation and arrays are always 64 elements if (get_quant_level(quant_level) <= 16) { - vint4 tab0(reinterpret_cast(qat.quant_to_unquant)); + vint4 tab0 = vint4::load(qat.quant_to_unquant); vint tab0p; vtable_prepare(tab0, tab0p); @@ -1056,8 +1056,8 @@ void compute_quantized_weights_for_decimation( } else { - vint4 tab0(reinterpret_cast(qat.quant_to_unquant)); - vint4 tab1(reinterpret_cast(qat.quant_to_unquant + 16)); + vint4 tab0 = vint4::load(qat.quant_to_unquant + 0); + vint4 tab1 = vint4::load(qat.quant_to_unquant + 16); vint tab0p, tab1p; vtable_prepare(tab0, tab1, tab0p, tab1p); diff --git a/3rdparty/astc-encoder/source/astcenc_image.cpp b/3rdparty/astc-encoder/source/astcenc_image.cpp index 9c0d672..b60d9cd 100644 --- a/3rdparty/astc-encoder/source/astcenc_image.cpp +++ b/3rdparty/astc-encoder/source/astcenc_image.cpp @@ -433,7 +433,7 @@ void store_image_block( vint data_rgbai = interleave_rgba8(data_ri, data_gi, data_bi, data_ai); vmask store_mask = vint::lane_id() < vint(used_texels); - store_lanes_masked(reinterpret_cast(data8_row), data_rgbai, store_mask); + store_lanes_masked(data8_row, data_rgbai, store_mask); data8_row += ASTCENC_SIMD_WIDTH * 4; idx += used_texels; diff --git a/3rdparty/astc-encoder/source/astcenc_internal.h b/3rdparty/astc-encoder/source/astcenc_internal.h index 63bbf8a..b1da41b 100644 --- a/3rdparty/astc-encoder/source/astcenc_internal.h +++ b/3rdparty/astc-encoder/source/astcenc_internal.h @@ -79,7 +79,7 @@ static constexpr unsigned int BLOCK_MAX_PARTITIONS { 4 }; /** @brief The number of partitionings, per partition count, suported by the ASTC format. */ static constexpr unsigned int BLOCK_MAX_PARTITIONINGS { 1024 }; -/** @brief The maximum number of weights used during partition selection for texel clustering. */ +/** @brief The maximum number of texels used during partition selection for texel clustering. */ static constexpr uint8_t BLOCK_MAX_KMEANS_TEXELS { 64 }; /** @brief The maximum number of weights a block can support. */ @@ -119,11 +119,9 @@ static constexpr unsigned int WEIGHTS_MAX_DECIMATION_MODES { 87 }; static constexpr float ERROR_CALC_DEFAULT { 1e30f }; /** - * @brief The minimum texel count for a block to use the one partition fast path. - * - * This setting skips 4x4 and 5x4 block sizes. + * @brief The minimum tuning setting threshold for the one partition fast path. */ -static constexpr unsigned int TUNE_MIN_TEXELS_MODE0_FASTPATH { 24 }; +static constexpr float TUNE_MIN_SEARCH_MODE0 { 0.85f }; /** * @brief The maximum number of candidate encodings tested for each encoding mode. @@ -137,7 +135,7 @@ static constexpr unsigned int TUNE_MAX_TRIAL_CANDIDATES { 8 }; * * This can be dynamically reduced by the compression quality preset. */ -static constexpr unsigned int TUNE_MAX_PARTITIONING_CANDIDATES { 32 }; +static constexpr unsigned int TUNE_MAX_PARTITIONING_CANDIDATES { 8 }; /** * @brief The maximum quant level using full angular endpoint search method. @@ -1025,13 +1023,13 @@ struct dt_init_working_buffers struct quant_and_transfer_table { /** @brief The unscrambled unquantized value. */ - int8_t quant_to_unquant[32]; + uint8_t quant_to_unquant[32]; /** @brief The scrambling order: scrambled_quant = map[unscrambled_quant]. */ - int8_t scramble_map[32]; + uint8_t scramble_map[32]; /** @brief The unscrambling order: unscrambled_unquant = map[scrambled_quant]. */ - int8_t unscramble_and_unquant_map[32]; + uint8_t unscramble_and_unquant_map[32]; /** * @brief A table of previous-and-next weights, indexed by the current unquantized value. @@ -1060,7 +1058,7 @@ static constexpr uint8_t SYM_BTYPE_NONCONST { 3 }; * @brief A symbolic representation of a compressed block. * * The symbolic representation stores the unpacked content of a single - * @c physical_compressed_block, in a form which is much easier to access for + * physical compressed block, in a form which is much easier to access for * the rest of the compressor code. */ struct symbolic_compressed_block @@ -1122,18 +1120,6 @@ struct symbolic_compressed_block } }; -/** - * @brief A physical representation of a compressed block. - * - * The physical representation stores the raw bytes of the format in memory. - */ -struct physical_compressed_block -{ - /** @brief The ASTC encoded data for a single block. */ - uint8_t data[16]; -}; - - /** * @brief Parameter structure for @c compute_pixel_region_variance(). * @@ -1848,6 +1834,34 @@ void unpack_color_endpoints( vint4& output0, vint4& output1); +/** + * @brief Unpack an LDR RGBA color that uses delta encoding. + * + * @param input0 The packed endpoint 0 color. + * @param input1 The packed endpoint 1 color deltas. + * @param[out] output0 The unpacked endpoint 0 color. + * @param[out] output1 The unpacked endpoint 1 color. + */ +void rgba_delta_unpack( + vint4 input0, + vint4 input1, + vint4& output0, + vint4& output1); + +/** + * @brief Unpack an LDR RGBA color that uses direct encoding. + * + * @param input0 The packed endpoint 0 color. + * @param input1 The packed endpoint 1 color. + * @param[out] output0 The unpacked endpoint 0 color. + * @param[out] output1 The unpacked endpoint 1 color. + */ +void rgba_unpack( + vint4 input0, + vint4 input1, + vint4& output0, + vint4& output1); + /** * @brief Unpack a set of quantized and decimated weights. * @@ -2007,7 +2021,7 @@ void compute_angular_endpoints_2planes( void compress_block( const astcenc_contexti& ctx, const image_block& blk, - physical_compressed_block& pcb, + uint8_t pcb[16], compression_working_buffers& tmpbuf); /** @@ -2100,12 +2114,12 @@ float compute_symbolic_block_difference_1plane_1partition( * * @param bsd The block size information. * @param scb The symbolic representation. - * @param[out] pcb The binary encoded data. + * @param[out] pcb The physical compressed block output. */ void symbolic_to_physical( const block_size_descriptor& bsd, const symbolic_compressed_block& scb, - physical_compressed_block& pcb); + uint8_t pcb[16]); /** * @brief Convert a binary physical encoding into a symbolic representation. @@ -2114,12 +2128,12 @@ void symbolic_to_physical( * flagged as an error block if the encoding is invalid. * * @param bsd The block size information. - * @param pcb The binary encoded data. + * @param pcb The physical compresesd block input. * @param[out] scb The output symbolic representation. */ void physical_to_symbolic( const block_size_descriptor& bsd, - const physical_compressed_block& pcb, + const uint8_t pcb[16], symbolic_compressed_block& scb); /* ============================================================================ @@ -2164,9 +2178,9 @@ template void aligned_free(T* ptr) { #if defined(_WIN32) - _aligned_free(reinterpret_cast(ptr)); + _aligned_free(ptr); #else - free(reinterpret_cast(ptr)); + free(ptr); #endif } diff --git a/3rdparty/astc-encoder/source/astcenc_symbolic_physical.cpp b/3rdparty/astc-encoder/source/astcenc_symbolic_physical.cpp index 49a8a15..c4da678 100644 --- a/3rdparty/astc-encoder/source/astcenc_symbolic_physical.cpp +++ b/3rdparty/astc-encoder/source/astcenc_symbolic_physical.cpp @@ -102,7 +102,7 @@ static inline void write_bits( void symbolic_to_physical( const block_size_descriptor& bsd, const symbolic_compressed_block& scb, - physical_compressed_block& pcb + uint8_t pcb[16] ) { assert(scb.block_type != SYM_BTYPE_ERROR); @@ -113,13 +113,13 @@ void symbolic_to_physical( static const uint8_t cbytes[8] { 0xFC, 0xFD, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF }; for (unsigned int i = 0; i < 8; i++) { - pcb.data[i] = cbytes[i]; + pcb[i] = cbytes[i]; } for (unsigned int i = 0; i < BLOCK_MAX_COMPONENTS; i++) { - pcb.data[2 * i + 8] = scb.constant_color[i] & 0xFF; - pcb.data[2 * i + 9] = (scb.constant_color[i] >> 8) & 0xFF; + pcb[2 * i + 8] = scb.constant_color[i] & 0xFF; + pcb[2 * i + 9] = (scb.constant_color[i] >> 8) & 0xFF; } return; @@ -132,13 +132,13 @@ void symbolic_to_physical( static const uint8_t cbytes[8] { 0xFC, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF }; for (unsigned int i = 0; i < 8; i++) { - pcb.data[i] = cbytes[i]; + pcb[i] = cbytes[i]; } for (unsigned int i = 0; i < BLOCK_MAX_COMPONENTS; i++) { - pcb.data[2 * i + 8] = scb.constant_color[i] & 0xFF; - pcb.data[2 * i + 9] = (scb.constant_color[i] >> 8) & 0xFF; + pcb[2 * i + 8] = scb.constant_color[i] & 0xFF; + pcb[2 * i + 9] = (scb.constant_color[i] >> 8) & 0xFF; } return; @@ -194,23 +194,23 @@ void symbolic_to_physical( for (int i = 0; i < 16; i++) { - pcb.data[i] = static_cast(bitrev8(weightbuf[15 - i])); + pcb[i] = static_cast(bitrev8(weightbuf[15 - i])); } - write_bits(scb.block_mode, 11, 0, pcb.data); - write_bits(partition_count - 1, 2, 11, pcb.data); + write_bits(scb.block_mode, 11, 0, pcb); + write_bits(partition_count - 1, 2, 11, pcb); int below_weights_pos = 128 - bits_for_weights; // Encode partition index and color endpoint types for blocks with 2+ partitions if (partition_count > 1) { - write_bits(scb.partition_index, 6, 13, pcb.data); - write_bits(scb.partition_index >> 6, PARTITION_INDEX_BITS - 6, 19, pcb.data); + write_bits(scb.partition_index, 6, 13, pcb); + write_bits(scb.partition_index >> 6, PARTITION_INDEX_BITS - 6, 19, pcb); if (scb.color_formats_matched) { - write_bits(scb.color_formats[0] << 2, 6, 13 + PARTITION_INDEX_BITS, pcb.data); + write_bits(scb.color_formats[0] << 2, 6, 13 + PARTITION_INDEX_BITS, pcb); } else { @@ -249,20 +249,20 @@ void symbolic_to_physical( int encoded_type_highpart = encoded_type >> 6; int encoded_type_highpart_size = (3 * partition_count) - 4; int encoded_type_highpart_pos = 128 - bits_for_weights - encoded_type_highpart_size; - write_bits(encoded_type_lowpart, 6, 13 + PARTITION_INDEX_BITS, pcb.data); - write_bits(encoded_type_highpart, encoded_type_highpart_size, encoded_type_highpart_pos, pcb.data); + write_bits(encoded_type_lowpart, 6, 13 + PARTITION_INDEX_BITS, pcb); + write_bits(encoded_type_highpart, encoded_type_highpart_size, encoded_type_highpart_pos, pcb); below_weights_pos -= encoded_type_highpart_size; } } else { - write_bits(scb.color_formats[0], 4, 13, pcb.data); + write_bits(scb.color_formats[0], 4, 13, pcb); } // In dual-plane mode, encode the color component of the second plane of weights if (is_dual_plane) { - write_bits(scb.plane2_component, 2, below_weights_pos - 2, pcb.data); + write_bits(scb.plane2_component, 2, below_weights_pos - 2, pcb); } // Encode the color components @@ -281,7 +281,7 @@ void symbolic_to_physical( valuecount_to_encode += vals; } - encode_ise(scb.get_color_quant_mode(), valuecount_to_encode, values_to_encode, pcb.data, + encode_ise(scb.get_color_quant_mode(), valuecount_to_encode, values_to_encode, pcb, scb.partition_count == 1 ? 17 : 19 + PARTITION_INDEX_BITS); } @@ -290,7 +290,7 @@ void symbolic_to_physical( /* See header for documentation. */ void physical_to_symbolic( const block_size_descriptor& bsd, - const physical_compressed_block& pcb, + const uint8_t pcb[16], symbolic_compressed_block& scb ) { uint8_t bswapped[16]; @@ -298,7 +298,7 @@ void physical_to_symbolic( scb.block_type = SYM_BTYPE_NONCONST; // Extract header fields - int block_mode = read_bits(11, 0, pcb.data); + int block_mode = read_bits(11, 0, pcb); if ((block_mode & 0x1FF) == 0x1FC) { // Constant color block @@ -316,24 +316,24 @@ void physical_to_symbolic( scb.partition_count = 0; for (int i = 0; i < 4; i++) { - scb.constant_color[i] = pcb.data[2 * i + 8] | (pcb.data[2 * i + 9] << 8); + scb.constant_color[i] = pcb[2 * i + 8] | (pcb[2 * i + 9] << 8); } // Additionally, check that the void-extent if (bsd.zdim == 1) { // 2D void-extent - int rsvbits = read_bits(2, 10, pcb.data); + int rsvbits = read_bits(2, 10, pcb); if (rsvbits != 3) { scb.block_type = SYM_BTYPE_ERROR; return; } - int vx_low_s = read_bits(8, 12, pcb.data) | (read_bits(5, 12 + 8, pcb.data) << 8); - int vx_high_s = read_bits(8, 25, pcb.data) | (read_bits(5, 25 + 8, pcb.data) << 8); - int vx_low_t = read_bits(8, 38, pcb.data) | (read_bits(5, 38 + 8, pcb.data) << 8); - int vx_high_t = read_bits(8, 51, pcb.data) | (read_bits(5, 51 + 8, pcb.data) << 8); + int vx_low_s = read_bits(8, 12, pcb) | (read_bits(5, 12 + 8, pcb) << 8); + int vx_high_s = read_bits(8, 25, pcb) | (read_bits(5, 25 + 8, pcb) << 8); + int vx_low_t = read_bits(8, 38, pcb) | (read_bits(5, 38 + 8, pcb) << 8); + int vx_high_t = read_bits(8, 51, pcb) | (read_bits(5, 51 + 8, pcb) << 8); int all_ones = vx_low_s == 0x1FFF && vx_high_s == 0x1FFF && vx_low_t == 0x1FFF && vx_high_t == 0x1FFF; @@ -346,12 +346,12 @@ void physical_to_symbolic( else { // 3D void-extent - int vx_low_s = read_bits(9, 10, pcb.data); - int vx_high_s = read_bits(9, 19, pcb.data); - int vx_low_t = read_bits(9, 28, pcb.data); - int vx_high_t = read_bits(9, 37, pcb.data); - int vx_low_p = read_bits(9, 46, pcb.data); - int vx_high_p = read_bits(9, 55, pcb.data); + int vx_low_s = read_bits(9, 10, pcb); + int vx_high_s = read_bits(9, 19, pcb); + int vx_low_t = read_bits(9, 28, pcb); + int vx_high_t = read_bits(9, 37, pcb); + int vx_low_p = read_bits(9, 46, pcb); + int vx_high_p = read_bits(9, 55, pcb); int all_ones = vx_low_s == 0x1FF && vx_high_s == 0x1FF && vx_low_t == 0x1FF && vx_high_t == 0x1FF && vx_low_p == 0x1FF && vx_high_p == 0x1FF; @@ -383,7 +383,7 @@ void physical_to_symbolic( int real_weight_count = is_dual_plane ? 2 * weight_count : weight_count; - int partition_count = read_bits(2, 11, pcb.data) + 1; + int partition_count = read_bits(2, 11, pcb) + 1; promise(partition_count > 0); scb.block_mode = static_cast(block_mode); @@ -391,7 +391,7 @@ void physical_to_symbolic( for (int i = 0; i < 16; i++) { - bswapped[i] = static_cast(bitrev8(pcb.data[15 - i])); + bswapped[i] = static_cast(bitrev8(pcb[15 - i])); } int bits_for_weights = get_ise_sequence_bitcount(real_weight_count, weight_quant_method); @@ -432,14 +432,15 @@ void physical_to_symbolic( int encoded_type_highpart_size = 0; if (partition_count == 1) { - color_formats[0] = read_bits(4, 13, pcb.data); + color_formats[0] = read_bits(4, 13, pcb); scb.partition_index = 0; } else { encoded_type_highpart_size = (3 * partition_count) - 4; below_weights_pos -= encoded_type_highpart_size; - int encoded_type = read_bits(6, 13 + PARTITION_INDEX_BITS, pcb.data) | (read_bits(encoded_type_highpart_size, below_weights_pos, pcb.data) << 6); + int encoded_type = read_bits(6, 13 + PARTITION_INDEX_BITS, pcb) | + (read_bits(encoded_type_highpart_size, below_weights_pos, pcb) << 6); int baseclass = encoded_type & 0x3; if (baseclass == 0) { @@ -469,7 +470,8 @@ void physical_to_symbolic( bitpos += 2; } } - scb.partition_index = static_cast(read_bits(6, 13, pcb.data) | (read_bits(PARTITION_INDEX_BITS - 6, 19, pcb.data) << 6)); + scb.partition_index = static_cast(read_bits(6, 13, pcb) | + (read_bits(PARTITION_INDEX_BITS - 6, 19, pcb) << 6)); } for (int i = 0; i < partition_count; i++) @@ -515,7 +517,7 @@ void physical_to_symbolic( scb.quant_mode = static_cast(color_quant_level); uint8_t values_to_decode[32]; - decode_ise(static_cast(color_quant_level), color_integer_count, pcb.data, + decode_ise(static_cast(color_quant_level), color_integer_count, pcb, values_to_decode, (partition_count == 1 ? 17 : 19 + PARTITION_INDEX_BITS)); int valuecount_to_decode = 0; @@ -534,6 +536,6 @@ void physical_to_symbolic( scb.plane2_component = -1; if (is_dual_plane) { - scb.plane2_component = static_cast(read_bits(2, below_weights_pos - 2, pcb.data)); + scb.plane2_component = static_cast(read_bits(2, below_weights_pos - 2, pcb)); } } diff --git a/3rdparty/astc-encoder/source/astcenc_vecmathlib_avx2_8.h b/3rdparty/astc-encoder/source/astcenc_vecmathlib_avx2_8.h index a785aca..72ed19f 100644 --- a/3rdparty/astc-encoder/source/astcenc_vecmathlib_avx2_8.h +++ b/3rdparty/astc-encoder/source/astcenc_vecmathlib_avx2_8.h @@ -241,6 +241,14 @@ struct vint8 return vint8(_mm256_broadcastd_epi32(a)); } + /** + * @brief Factory that returns a vector loaded from unaligned memory. + */ + static ASTCENC_SIMD_INLINE vint8 load(const uint8_t* p) + { + return vint8(_mm256_lddqu_si256(reinterpret_cast(p))); + } + /** * @brief Factory that returns a vector loaded from 32B aligned memory. */ @@ -1000,7 +1008,7 @@ ASTCENC_SIMD_INLINE vint8 float_to_int(vfloat8 a) */ ASTCENC_SIMD_INLINE vint8 float_to_int_rtn(vfloat8 a) { - a = round(a); + a = a + vfloat8(0.5f); return vint8(_mm256_cvttps_epi32(a.m)); } @@ -1152,9 +1160,9 @@ ASTCENC_SIMD_INLINE vint8 interleave_rgba8(vint8 r, vint8 g, vint8 b, vint8 a) * * All masked lanes must be at the end of vector, after all non-masked lanes. */ -ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint8 data, vmask8 mask) +ASTCENC_SIMD_INLINE void store_lanes_masked(uint8_t* base, vint8 data, vmask8 mask) { - _mm256_maskstore_epi32(base, _mm256_castps_si256(mask.m), data.m); + _mm256_maskstore_epi32(reinterpret_cast(base), _mm256_castps_si256(mask.m), data.m); } /** diff --git a/3rdparty/astc-encoder/source/astcenc_vecmathlib_neon_4.h b/3rdparty/astc-encoder/source/astcenc_vecmathlib_neon_4.h index e742eae..c5ad872 100644 --- a/3rdparty/astc-encoder/source/astcenc_vecmathlib_neon_4.h +++ b/3rdparty/astc-encoder/source/astcenc_vecmathlib_neon_4.h @@ -1,6 +1,6 @@ // SPDX-License-Identifier: Apache-2.0 // ---------------------------------------------------------------------------- -// Copyright 2019-2022 Arm Limited +// Copyright 2019-2023 Arm Limited // // Licensed under the Apache License, Version 2.0 (the "License"); you may not // use this file except in compliance with the License. You may obtain a copy @@ -38,6 +38,7 @@ #endif #include +#include // ============================================================================ // vfloat4 data type @@ -269,6 +270,16 @@ struct vint4 return vint4(*p); } + /** + * @brief Factory that returns a vector loaded from unaligned memory. + */ + static ASTCENC_SIMD_INLINE vint4 load(const uint8_t* p) + { + vint4 data; + std::memcpy(&data.m, p, 4 * sizeof(int)); + return data; + } + /** * @brief Factory that returns a vector loaded from 16B aligned memory. */ @@ -584,6 +595,14 @@ ASTCENC_SIMD_INLINE void store(vint4 a, int* p) vst1q_s32(p, a.m); } +/** + * @brief Store a vector to an unaligned memory address. + */ +ASTCENC_SIMD_INLINE void store(vint4 a, uint8_t* p) +{ + std::memcpy(p, &a.m, sizeof(int) * 4); +} + /** * @brief Store lowest N (vector width) bytes into an unaligned address. */ @@ -849,7 +868,7 @@ ASTCENC_SIMD_INLINE vint4 float_to_int(vfloat4 a) */ ASTCENC_SIMD_INLINE vint4 float_to_int_rtn(vfloat4 a) { - a = round(a); + a = a + vfloat4(0.5f); return vint4(vcvtq_s32_f32(a.m)); } @@ -1027,31 +1046,39 @@ ASTCENC_SIMD_INLINE vint4 interleave_rgba8(vint4 r, vint4 g, vint4 b, vint4 a) return r + lsl<8>(g) + lsl<16>(b) + lsl<24>(a); } +/** + * @brief Store a single vector lane to an unaligned address. + */ +ASTCENC_SIMD_INLINE void store_lane(uint8_t* base, int data) +{ + std::memcpy(base, &data, sizeof(int)); +} + /** * @brief Store a vector, skipping masked lanes. * * All masked lanes must be at the end of vector, after all non-masked lanes. */ -ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint4 data, vmask4 mask) +ASTCENC_SIMD_INLINE void store_lanes_masked(uint8_t* base, vint4 data, vmask4 mask) { if (mask.lane<3>()) { store(data, base); } - else if (mask.lane<2>()) + else if (mask.lane<2>() != 0.0f) { - base[0] = data.lane<0>(); - base[1] = data.lane<1>(); - base[2] = data.lane<2>(); + store_lane(base + 0, data.lane<0>()); + store_lane(base + 4, data.lane<1>()); + store_lane(base + 8, data.lane<2>()); } - else if (mask.lane<1>()) + else if (mask.lane<1>() != 0.0f) { - base[0] = data.lane<0>(); - base[1] = data.lane<1>(); + store_lane(base + 0, data.lane<0>()); + store_lane(base + 4, data.lane<1>()); } - else if (mask.lane<0>()) + else if (mask.lane<0>() != 0.0f) { - base[0] = data.lane<0>(); + store_lane(base + 0, data.lane<0>()); } } diff --git a/3rdparty/astc-encoder/source/astcenc_vecmathlib_none_4.h b/3rdparty/astc-encoder/source/astcenc_vecmathlib_none_4.h index d9b52be..6dbb659 100644 --- a/3rdparty/astc-encoder/source/astcenc_vecmathlib_none_4.h +++ b/3rdparty/astc-encoder/source/astcenc_vecmathlib_none_4.h @@ -275,6 +275,16 @@ struct vint4 return vint4(*p); } + /** + * @brief Factory that returns a vector loaded from unaligned memory. + */ + static ASTCENC_SIMD_INLINE vint4 load(const uint8_t* p) + { + vint4 data; + std::memcpy(&data.m, p, 4 * sizeof(int)); + return data; + } + /** * @brief Factory that returns a vector loaded from 16B aligned memory. */ @@ -644,13 +654,20 @@ ASTCENC_SIMD_INLINE void store(vint4 a, int* p) p[3] = a.m[3]; } +/** + * @brief Store a vector to an unaligned memory address. + */ +ASTCENC_SIMD_INLINE void store(vint4 a, uint8_t* p) +{ + std::memcpy(p, a.m, sizeof(int) * 4); +} + /** * @brief Store lowest N (vector width) bytes into an unaligned address. */ ASTCENC_SIMD_INLINE void store_nbytes(vint4 a, uint8_t* p) { - int* pi = reinterpret_cast(p); - *pi = a.m[0]; + std::memcpy(p, a.m, sizeof(uint8_t) * 4); } /** @@ -963,10 +980,11 @@ ASTCENC_SIMD_INLINE vint4 float_to_int(vfloat4 a) */ ASTCENC_SIMD_INLINE vint4 float_to_int_rtn(vfloat4 a) { - return vint4(static_cast(a.m[0] + 0.5f), - static_cast(a.m[1] + 0.5f), - static_cast(a.m[2] + 0.5f), - static_cast(a.m[3] + 0.5f)); + a = a + vfloat4(0.5f); + return vint4(static_cast(a.m[0]), + static_cast(a.m[1]), + static_cast(a.m[2]), + static_cast(a.m[3])); } /** @@ -1030,7 +1048,7 @@ ASTCENC_SIMD_INLINE float float16_to_float(uint16_t a) ASTCENC_SIMD_INLINE vint4 float_as_int(vfloat4 a) { vint4 r; - memcpy(r.m, a.m, 4 * 4); + std::memcpy(r.m, a.m, 4 * 4); return r; } @@ -1044,7 +1062,7 @@ ASTCENC_SIMD_INLINE vint4 float_as_int(vfloat4 a) ASTCENC_SIMD_INLINE vfloat4 int_as_float(vint4 a) { vfloat4 r; - memcpy(r.m, a.m, 4 * 4); + std::memcpy(r.m, a.m, 4 * 4); return r; } @@ -1079,12 +1097,13 @@ ASTCENC_SIMD_INLINE void vtable_prepare( } /** - * @brief Perform an 8-bit 32-entry table lookup, with 32-bit indexes. + * @brief Perform an 8-bit 16-entry table lookup, with 32-bit indexes. */ ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 idx) { uint8_t table[16]; - storea(t0, reinterpret_cast(table + 0)); + + std::memcpy(table + 0, t0.m, 4 * sizeof(int)); return vint4(table[idx.lane<0>()], table[idx.lane<1>()], @@ -1099,8 +1118,9 @@ ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 idx) ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 idx) { uint8_t table[32]; - storea(t0, reinterpret_cast(table + 0)); - storea(t1, reinterpret_cast(table + 16)); + + std::memcpy(table + 0, t0.m, 4 * sizeof(int)); + std::memcpy(table + 16, t1.m, 4 * sizeof(int)); return vint4(table[idx.lane<0>()], table[idx.lane<1>()], @@ -1114,10 +1134,11 @@ ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 idx) ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 t2, vint4 t3, vint4 idx) { uint8_t table[64]; - storea(t0, reinterpret_cast(table + 0)); - storea(t1, reinterpret_cast(table + 16)); - storea(t2, reinterpret_cast(table + 32)); - storea(t3, reinterpret_cast(table + 48)); + + std::memcpy(table + 0, t0.m, 4 * sizeof(int)); + std::memcpy(table + 16, t1.m, 4 * sizeof(int)); + std::memcpy(table + 32, t2.m, 4 * sizeof(int)); + std::memcpy(table + 48, t3.m, 4 * sizeof(int)); return vint4(table[idx.lane<0>()], table[idx.lane<1>()], @@ -1138,12 +1159,21 @@ ASTCENC_SIMD_INLINE vint4 interleave_rgba8(vint4 r, vint4 g, vint4 b, vint4 a) return r + lsl<8>(g) + lsl<16>(b) + lsl<24>(a); } +/** + * @brief Store a single vector lane to an unaligned address. + */ +ASTCENC_SIMD_INLINE void store_lane(uint8_t* base, int data) +{ + std::memcpy(base, &data, sizeof(int)); +} + /** * @brief Store a vector, skipping masked lanes. * * All masked lanes must be at the end of vector, after all non-masked lanes. + * Input is a byte array of at least 4 bytes per unmasked entry. */ -ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint4 data, vmask4 mask) +ASTCENC_SIMD_INLINE void store_lanes_masked(uint8_t* base, vint4 data, vmask4 mask) { if (mask.m[3]) { @@ -1151,18 +1181,18 @@ ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint4 data, vmask4 mask) } else if (mask.m[2]) { - base[0] = data.lane<0>(); - base[1] = data.lane<1>(); - base[2] = data.lane<2>(); + store_lane(base + 0, data.lane<0>()); + store_lane(base + 4, data.lane<1>()); + store_lane(base + 8, data.lane<2>()); } else if (mask.m[1]) { - base[0] = data.lane<0>(); - base[1] = data.lane<1>(); + store_lane(base + 0, data.lane<0>()); + store_lane(base + 4, data.lane<1>()); } else if (mask.m[0]) { - base[0] = data.lane<0>(); + store_lane(base + 0, data.lane<0>()); } } diff --git a/3rdparty/astc-encoder/source/astcenc_vecmathlib_sse_4.h b/3rdparty/astc-encoder/source/astcenc_vecmathlib_sse_4.h index b998107..4dd58d2 100644 --- a/3rdparty/astc-encoder/source/astcenc_vecmathlib_sse_4.h +++ b/3rdparty/astc-encoder/source/astcenc_vecmathlib_sse_4.h @@ -39,6 +39,7 @@ #endif #include +#include // ============================================================================ // vfloat4 data type @@ -292,6 +293,18 @@ struct vint4 return vint4(*p); } + /** + * @brief Factory that returns a vector loaded from unaligned memory. + */ + static ASTCENC_SIMD_INLINE vint4 load(const uint8_t* p) + { +#if ASTCENC_SSE >= 41 + return vint4(_mm_lddqu_si128(reinterpret_cast(p))); +#else + return vint4(_mm_loadu_si128(reinterpret_cast(p))); +#endif + } + /** * @brief Factory that returns a vector loaded from 16B aligned memory. */ @@ -633,6 +646,14 @@ ASTCENC_SIMD_INLINE void store(vint4 a, int* p) _mm_storeu_ps(reinterpret_cast(p), _mm_castsi128_ps(a.m)); } +/** + * @brief Store a vector to an unaligned memory address. + */ +ASTCENC_SIMD_INLINE void store(vint4 a, uint8_t* p) +{ + std::memcpy(p, &a.m, sizeof(int) * 4); +} + /** * @brief Store lowest N (vector width) bytes into an unaligned address. */ @@ -934,7 +955,7 @@ ASTCENC_SIMD_INLINE vint4 float_to_int(vfloat4 a) */ ASTCENC_SIMD_INLINE vint4 float_to_int_rtn(vfloat4 a) { - a = round(a); + a = a + vfloat4(0.5f); return vint4(_mm_cvttps_epi32(a.m)); } @@ -1087,8 +1108,9 @@ ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 idx) __m128i result = _mm_shuffle_epi8(t0.m, idxx); return vint4(result); #else - alignas(ASTCENC_VECALIGN) uint8_t table[16]; - storea(t0, reinterpret_cast(table + 0)); + uint8_t table[16]; + + std::memcpy(table + 0, &t0.m, 4 * sizeof(int)); return vint4(table[idx.lane<0>()], table[idx.lane<1>()], @@ -1114,9 +1136,10 @@ ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 idx) return vint4(result); #else - alignas(ASTCENC_VECALIGN) uint8_t table[32]; - storea(t0, reinterpret_cast(table + 0)); - storea(t1, reinterpret_cast(table + 16)); + uint8_t table[32]; + + std::memcpy(table + 0, &t0.m, 4 * sizeof(int)); + std::memcpy(table + 16, &t1.m, 4 * sizeof(int)); return vint4(table[idx.lane<0>()], table[idx.lane<1>()], @@ -1150,11 +1173,12 @@ ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 t2, vint4 t3 return vint4(result); #else - alignas(ASTCENC_VECALIGN) uint8_t table[64]; - storea(t0, reinterpret_cast(table + 0)); - storea(t1, reinterpret_cast(table + 16)); - storea(t2, reinterpret_cast(table + 32)); - storea(t3, reinterpret_cast(table + 48)); + uint8_t table[64]; + + std::memcpy(table + 0, &t0.m, 4 * sizeof(int)); + std::memcpy(table + 16, &t1.m, 4 * sizeof(int)); + std::memcpy(table + 32, &t2.m, 4 * sizeof(int)); + std::memcpy(table + 48, &t3.m, 4 * sizeof(int)); return vint4(table[idx.lane<0>()], table[idx.lane<1>()], @@ -1190,15 +1214,23 @@ ASTCENC_SIMD_INLINE vint4 interleave_rgba8(vint4 r, vint4 g, vint4 b, vint4 a) #endif } +/** + * @brief Store a single vector lane to an unaligned address. + */ +ASTCENC_SIMD_INLINE void store_lane(uint8_t* base, int data) +{ + std::memcpy(base, &data, sizeof(int)); +} + /** * @brief Store a vector, skipping masked lanes. * * All masked lanes must be at the end of vector, after all non-masked lanes. */ -ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint4 data, vmask4 mask) +ASTCENC_SIMD_INLINE void store_lanes_masked(uint8_t* base, vint4 data, vmask4 mask) { #if ASTCENC_AVX >= 2 - _mm_maskstore_epi32(base, _mm_castps_si128(mask.m), data.m); + _mm_maskstore_epi32(reinterpret_cast(base), _mm_castps_si128(mask.m), data.m); #else // Note - we cannot use _mm_maskmoveu_si128 as the underlying hardware doesn't guarantee // fault suppression on masked lanes so we can get page faults at the end of an image. @@ -1208,18 +1240,18 @@ ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint4 data, vmask4 mask) } else if (mask.lane<2>() != 0.0f) { - base[0] = data.lane<0>(); - base[1] = data.lane<1>(); - base[2] = data.lane<2>(); + store_lane(base + 0, data.lane<0>()); + store_lane(base + 4, data.lane<1>()); + store_lane(base + 8, data.lane<2>()); } else if (mask.lane<1>() != 0.0f) { - base[0] = data.lane<0>(); - base[1] = data.lane<1>(); + store_lane(base + 0, data.lane<0>()); + store_lane(base + 4, data.lane<1>()); } else if (mask.lane<0>() != 0.0f) { - base[0] = data.lane<0>(); + store_lane(base + 0, data.lane<0>()); } #endif }