Add Arm SVE 8-wide (256b) implementation

CMakeLists.txt

@@ -38,6 +38,7 @@ include(CTest)
 option(ASTCENC_ISA_AVX2 "Enable astcenc builds for AVX2 SIMD")
 option(ASTCENC_ISA_SSE41 "Enable astcenc builds for SSE4.1 SIMD")
 option(ASTCENC_ISA_SSE2 "Enable astcenc builds for SSE2 SIMD")
+option(ASTCENC_ISA_SVE_256 "Enable astcenc builds for 256-bit SVE SIMD")
 option(ASTCENC_ISA_NEON "Enable astcenc builds for NEON SIMD")
 option(ASTCENC_ISA_NONE "Enable astcenc builds for no SIMD")
 option(ASTCENC_ISA_NATIVE "Enable astcenc builds for native SIMD")
@@ -86,7 +87,7 @@ endforeach()
 
 # Count options which MUST be arm64
 set(ASTCENC_ARM64_ISA_COUNT 0)
-set(ASTCENC_CONFIGS ${ASTCENC_ISA_NEON})
+set(ASTCENC_CONFIGS ${ASTCENC_ISA_NEON} ${ASTCENC_ISA_SVE_256})
 foreach(ASTCENC_CONFIG ${ASTCENC_CONFIGS})
     if(${ASTCENC_CONFIG})
         math(EXPR ASTCENC_ARM64_ISA_COUNT "${ASTCENC_ARM64_ISA_COUNT} + 1")
@@ -117,22 +118,28 @@ if("${ASTCENC_BLOCK_MAX_TEXELS}")
      message(STATUS "  Max block texels - ${ASTCENC_BLOCK_MAX_TEXELS}")
 endif()
 
-printopt("AVX2 backend   " ${ASTCENC_ISA_AVX2})
-printopt("SSE4.1 backend " ${ASTCENC_ISA_SSE41})
-printopt("SSE2 backend   " ${ASTCENC_ISA_SSE2})
-printopt("NEON backend   " ${ASTCENC_ISA_NEON})
-printopt("NONE backend   " ${ASTCENC_ISA_NONE})
-printopt("NATIVE backend " ${ASTCENC_ISA_NATIVE})
+message(STATUS "Arm backend options")
+printopt("SVE 256b backend   " ${ASTCENC_ISA_SVE_256})
+printopt("NEON backend       " ${ASTCENC_ISA_NEON})
+message(STATUS "x86-64 backend options")
+printopt("AVX2 backend       " ${ASTCENC_ISA_AVX2})
+printopt("SSE4.1 backend     " ${ASTCENC_ISA_SSE41})
+printopt("SSE2 backend       " ${ASTCENC_ISA_SSE2})
+message(STATUS "Agnostic backend options")
+printopt("NONE backend       " ${ASTCENC_ISA_NONE})
+printopt("NATIVE backend     " ${ASTCENC_ISA_NATIVE})
+message(STATUS "Build options")
 if("${CMAKE_SYSTEM_NAME}" STREQUAL "Darwin")
-    printopt("Universal bin  " ${ASTCENC_UNIVERSAL_BUILD})
+    printopt("Universal bin      " ${ASTCENC_UNIVERSAL_BUILD})
 endif()
-printopt("Invariance     " ${ASTCENC_INVARIANCE})
-printopt("Shared libs    " ${ASTCENC_SHAREDLIB})
-printopt("Decompressor   " ${ASTCENC_DECOMPRESSOR})
-printopt("Diagnostics    " ${ASTCENC_DIAGNOSTICS})
-printopt("ASAN           " ${ASTCENC_ASAN})
-printopt("UBSAN          " ${ASTCENC_UBSAN})
-printopt("Unit tests     " ${ASTCENC_UNITTEST})
+printopt("Invariance         " ${ASTCENC_INVARIANCE})
+printopt("Shared libs        " ${ASTCENC_SHAREDLIB})
+printopt("Decompressor       " ${ASTCENC_DECOMPRESSOR})
+message(STATUS "Developer options")
+printopt("Diagnostics        " ${ASTCENC_DIAGNOSTICS})
+printopt("ASAN               " ${ASTCENC_ASAN})
+printopt("UBSAN              " ${ASTCENC_UBSAN})
+printopt("Unit tests         " ${ASTCENC_UNITTEST})
 
 # Subcomponents
 add_subdirectory(Source)

Source/CMakeLists.txt

@@ -27,8 +27,8 @@ else()
     set(ASTCENC_CODEC enc)
 endif()
 
-set(ASTCENC_ARTIFACTS native none neon avx2 sse4.1 sse2)
-set(ASTCENC_CONFIGS ${ASTCENC_ISA_NATIVE} ${ASTCENC_ISA_NONE} ${ASTCENC_ISA_NEON} ${ASTCENC_ISA_AVX2} ${ASTCENC_ISA_SSE41} ${ASTCENC_ISA_SSE2})
+set(ASTCENC_ARTIFACTS native none sve_256 neon avx2 sse4.1 sse2)
+set(ASTCENC_CONFIGS ${ASTCENC_ISA_NATIVE} ${ASTCENC_ISA_NONE} ${ASTCENC_ISA_SVE_256} ${ASTCENC_ISA_NEON} ${ASTCENC_ISA_AVX2} ${ASTCENC_ISA_SSE41} ${ASTCENC_ISA_SSE2})
 list(LENGTH ASTCENC_ARTIFACTS ASTCENC_ARTIFACTS_LEN)
 math(EXPR ASTCENC_ARTIFACTS_LEN "${ASTCENC_ARTIFACTS_LEN} - 1")
 
@@ -38,7 +38,9 @@ foreach(INDEX RANGE ${ASTCENC_ARTIFACTS_LEN})
     if(${ASTCENC_CONFIG})
         set(ASTCENC_ISA_SIMD ${ASTCENC_ARTIFACT})
 
-        if(${ASTCENC_ISA_SIMD} MATCHES "neon")
+        if(${ASTCENC_ISA_SIMD} MATCHES "sve_256")
+           # Not suported on macOS
+        elseif(${ASTCENC_ISA_SIMD} MATCHES "neon")
            set(CMAKE_OSX_ARCHITECTURES arm64)
         elseif(${ASTCENC_ISA_SIMD} MATCHES "sse2")
            set(CMAKE_OSX_ARCHITECTURES x86_64)

Source/UnitTest/CMakeLists.txt

@@ -15,8 +15,8 @@
 #  under the License.
 #  ----------------------------------------------------------------------------
 
-set(ASTCENC_ARTIFACTS native none neon avx2 sse4.1 sse2)
-set(ASTCENC_CONFIGS ${ASTCENC_ISA_NATIVE} ${ASTCENC_ISA_NONE} ${ASTCENC_ISA_NEON} ${ASTCENC_ISA_AVX2} ${ASTCENC_ISA_SSE41} ${ASTCENC_ISA_SSE2})
+set(ASTCENC_ARTIFACTS native none sve_256 neon avx2 sse4.1 sse2)
+set(ASTCENC_CONFIGS ${ASTCENC_ISA_NATIVE} ${ASTCENC_ISA_NONE} ${ASTCENC_ISA_SVE_256} ${ASTCENC_ISA_NEON} ${ASTCENC_ISA_AVX2} ${ASTCENC_ISA_SSE41} ${ASTCENC_ISA_SSE2})
 list(LENGTH ASTCENC_ARTIFACTS ASTCENC_ARTIFACTS_LEN)
 math(EXPR ASTCENC_ARTIFACTS_LEN "${ASTCENC_ARTIFACTS_LEN} - 1")
 
@@ -26,7 +26,9 @@ foreach(INDEX RANGE ${ASTCENC_ARTIFACTS_LEN})
     if(${ASTCENC_CONFIG})
         set(ASTCENC_ISA_SIMD ${ASTCENC_ARTIFACT})
 
-        if(${ASTCENC_ISA_SIMD} MATCHES "neon")
+        if(${ASTCENC_ISA_SIMD} MATCHES "sve_256")
+           # Not supported on macOS
+        elseif(${ASTCENC_ISA_SIMD} MATCHES "neon")
            set(CMAKE_OSX_ARCHITECTURES arm64)
         elseif(${ASTCENC_ISA_SIMD} MATCHES "sse2")
            set(CMAKE_OSX_ARCHITECTURES x86_64)

Source/UnitTest/cmake_core.cmake

@@ -1,6 +1,6 @@
 #  SPDX-License-Identifier: Apache-2.0
 #  ----------------------------------------------------------------------------
-#  Copyright 2020-2023 Arm Limited
+#  Copyright 2020-2024 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
@@ -72,6 +72,7 @@ if(${ASTCENC_ISA_SIMD} MATCHES "none")
     target_compile_definitions(${ASTCENC_TEST}
         PRIVATE
             ASTCENC_NEON=0
+            ASTCENC_SVE=0
             ASTCENC_SSE=0
             ASTCENC_AVX=0
             ASTCENC_POPCNT=0
@@ -81,15 +82,32 @@ elseif(${ASTCENC_ISA_SIMD} MATCHES "neon")
     target_compile_definitions(${ASTCENC_TEST}
         PRIVATE
             ASTCENC_NEON=1
+            ASTCENC_SVE=0
             ASTCENC_SSE=0
             ASTCENC_AVX=0
             ASTCENC_POPCNT=0
             ASTCENC_F16C=0)
 
+elseif(${ASTCENC_ISA_SIMD} MATCHES "sve_256")
+    target_compile_definitions(${ASTCENC_TEST}
+        PRIVATE
+            ASTCENC_NEON=1
+            ASTCENC_SVE=8
+            ASTCENC_SSE=0
+            ASTCENC_AVX=0
+            ASTCENC_POPCNT=0
+            ASTCENC_F16C=0)
+
+    # Enable SVE
+    target_compile_options(${ASTCENC_TEST}
+        PRIVATE
+            -march=armv8-a+sve -msve-vector-bits=256)
+
 elseif(${ASTCENC_ISA_SIMD} MATCHES "sse2")
     target_compile_definitions(${ASTCENC_TEST}
         PRIVATE
             ASTCENC_NEON=0
+            ASTCENC_SVE=0
             ASTCENC_SSE=20
             ASTCENC_AVX=0
             ASTCENC_POPCNT=0
@@ -103,6 +121,7 @@ elseif(${ASTCENC_ISA_SIMD} MATCHES "sse4.1")
     target_compile_definitions(${ASTCENC_TEST}
         PRIVATE
             ASTCENC_NEON=0
+            ASTCENC_SVE=0
             ASTCENC_SSE=41
             ASTCENC_AVX=0
             ASTCENC_POPCNT=1
@@ -116,6 +135,7 @@ elseif(${ASTCENC_ISA_SIMD} MATCHES "avx2")
     target_compile_definitions(${ASTCENC_TEST}
         PRIVATE
             ASTCENC_NEON=0
+            ASTCENC_SVE=0
             ASTCENC_SSE=41
             ASTCENC_AVX=2
             ASTCENC_POPCNT=1

Source/UnitTest/test_simd.cpp

@@ -194,6 +194,7 @@ TEST(vfloat, ChangeSign)
 	vfloat4 a(-1.0f,  1.0f, -3.12f, 3.12f);
 	vfloat4 b(-1.0f, -1.0f,  3.12f, 3.12f);
 	vfloat4 r = change_sign(a, b);
+
 	EXPECT_EQ(r.lane<0>(),  1.0f);
 	EXPECT_EQ(r.lane<1>(), -1.0f);
 	EXPECT_EQ(r.lane<2>(), -3.12f);
@@ -205,6 +206,7 @@ TEST(vfloat, Atan)
 {
 	vfloat4 a(-0.15f, 0.0f, 0.9f, 2.1f);
 	vfloat4 r = atan(a);
+
 	EXPECT_NEAR(r.lane<0>(), -0.149061f, 0.005f);
 	EXPECT_NEAR(r.lane<1>(),  0.000000f, 0.005f);
 	EXPECT_NEAR(r.lane<2>(),  0.733616f, 0.005f);
@@ -217,6 +219,7 @@ TEST(vfloat, Atan2)
 	vfloat4 a(-0.15f, 0.0f, 0.9f, 2.1f);
 	vfloat4 b(1.15f, -3.0f, -0.9f, 1.1f);
 	vfloat4 r = atan2(a, b);
+
 	EXPECT_NEAR(r.lane<0>(), -0.129816f, 0.005f);
 	EXPECT_NEAR(r.lane<1>(),  3.141592f, 0.005f);
 	EXPECT_NEAR(r.lane<2>(),  2.360342f, 0.005f);
@@ -909,31 +912,6 @@ TEST(vfloat4, select)
 	EXPECT_EQ(r2.lane<3>(), 4.0f);
 }
 
-/** @brief Test vfloat4 select MSB only. */
-TEST(vfloat4, select_msb)
-{
-	int msb_set = static_cast<int>(0x80000000);
-	vint4 msb(msb_set, 0, msb_set, 0);
-	vmask4 cond(msb.m);
-
-	vfloat4 a(1.0f, 3.0f, 3.0f, 1.0f);
-	vfloat4 b(4.0f, 2.0f, 2.0f, 4.0f);
-
-	// Select in one direction
-	vfloat4 r1 = select_msb(a, b, cond);
-	EXPECT_EQ(r1.lane<0>(), 4.0f);
-	EXPECT_EQ(r1.lane<1>(), 3.0f);
-	EXPECT_EQ(r1.lane<2>(), 2.0f);
-	EXPECT_EQ(r1.lane<3>(), 1.0f);
-
-	// Select in the other
-	vfloat4 r2 = select_msb(b, a, cond);
-	EXPECT_EQ(r2.lane<0>(), 1.0f);
-	EXPECT_EQ(r2.lane<1>(), 2.0f);
-	EXPECT_EQ(r2.lane<2>(), 3.0f);
-	EXPECT_EQ(r2.lane<3>(), 4.0f);
-}
-
 /** @brief Test vfloat4 gatherf. */
 TEST(vfloat4, gatherf)
 {
@@ -1839,12 +1817,17 @@ TEST(vint4, store_lanes_masked_unaligned)
 	EXPECT_TRUE(all(result3v == expect3v));
 }
 
-/** @brief Test vint4 pack_low_bytes. */
-TEST(vint4, pack_low_bytes)
+/** @brief Test vint4 pack_and_store_low_bytes. */
+TEST(vint4, pack_and_store_low_bytes)
 {
 	vint4 a(1, 2, 3, 4);
-	vint4 r = pack_low_bytes(a);
-	EXPECT_EQ(r.lane<0>(), (4 << 24) | (3 << 16) | (2  << 8) | (1 << 0));
+	uint8_t bytes[4] { 0 };
+	pack_and_store_low_bytes(a, bytes);
+
+	EXPECT_EQ(bytes[0], 1);
+	EXPECT_EQ(bytes[1], 2);
+	EXPECT_EQ(bytes[2], 3);
+	EXPECT_EQ(bytes[3], 4);
 }
 
 /** @brief Test vint4 select. */
@@ -2711,46 +2694,6 @@ TEST(vfloat8, select)
 	EXPECT_EQ(ra[7], 4.0f);
 }
 
-/** @brief Test vfloat8 select MSB only. */
-TEST(vfloat8, select_msb)
-{
-	int msb_set = static_cast<int>(0x80000000);
-	vint8 msb = vint8_lit(msb_set, 0, msb_set, 0, msb_set, 0, msb_set, 0);
-	vmask8 cond(msb.m);
-
-	vfloat8 a = vfloat8_lit(1.0f, 3.0f, 3.0f, 1.0f, 1.0f, 3.0f, 3.0f, 1.0f);
-	vfloat8 b = vfloat8_lit(4.0f, 2.0f, 2.0f, 4.0f, 4.0f, 2.0f, 2.0f, 4.0f);
-
-	// Select in one direction
-	vfloat8 r1 = select(a, b, cond);
-
-	alignas(32) float ra[8];
-	storea(r1, ra);
-
-	EXPECT_EQ(ra[0], 4.0f);
-	EXPECT_EQ(ra[1], 3.0f);
-	EXPECT_EQ(ra[2], 2.0f);
-	EXPECT_EQ(ra[3], 1.0f);
-	EXPECT_EQ(ra[4], 4.0f);
-	EXPECT_EQ(ra[5], 3.0f);
-	EXPECT_EQ(ra[6], 2.0f);
-	EXPECT_EQ(ra[7], 1.0f);
-
-	// Select in the other
-	vfloat8 r2 = select(b, a, cond);
-
-	storea(r2, ra);
-
-	EXPECT_EQ(ra[0], 1.0f);
-	EXPECT_EQ(ra[1], 2.0f);
-	EXPECT_EQ(ra[2], 3.0f);
-	EXPECT_EQ(ra[3], 4.0f);
-	EXPECT_EQ(ra[4], 1.0f);
-	EXPECT_EQ(ra[5], 2.0f);
-	EXPECT_EQ(ra[6], 3.0f);
-	EXPECT_EQ(ra[7], 4.0f);
-}
-
 /** @brief Test vfloat8 gatherf. */
 TEST(vfloat8, gatherf)
 {
@@ -3583,17 +3526,22 @@ TEST(vint8, store_lanes_masked_unaligned)
 	EXPECT_TRUE(all(result3v == expect3v));
 }
 
-/** @brief Test vint8 pack_low_bytes. */
-TEST(vint8, pack_low_bytes)
+/** @brief Test vint8 pack_and_store_low_bytes. */
+TEST(vint8, pack_and_store_low_bytes)
 {
 	vint8 a = vint8_lit(1, 2, 3, 4, 2, 3, 4, 5);
-	vint8 r = pack_low_bytes(a);
-
-	alignas(32) int ra[8];
-	store(r, ra);
-
-	EXPECT_EQ(ra[0], (4 << 24) | (3 << 16) | (2  << 8) | (1 << 0));
-	EXPECT_EQ(ra[1], (5 << 24) | (4 << 16) | (3  << 8) | (2 << 0));
+	uint8_t bytes[8] { 0 };
+
+	pack_and_store_low_bytes(a, bytes);
+
+	EXPECT_EQ(bytes[0], 1);
+	EXPECT_EQ(bytes[1], 2);
+	EXPECT_EQ(bytes[2], 3);
+	EXPECT_EQ(bytes[3], 4);
+	EXPECT_EQ(bytes[4], 2);
+	EXPECT_EQ(bytes[5], 3);
+	EXPECT_EQ(bytes[6], 4);
+	EXPECT_EQ(bytes[7], 5);
 }
 
 /** @brief Test vint8 select. */

Source/astcenc_ideal_endpoints_and_weights.cpp

@@ -1050,8 +1050,7 @@ void compute_quantized_weights_for_decimation(
 
 			// Invert the weight-scaling that was done initially
 			storea(ixl * rscalev + low_boundv, weight_set_out + i);
-			vint scn = pack_low_bytes(weight);
-			store_nbytes(scn, quantized_weight_set + i);
+			pack_and_store_low_bytes(weight, quantized_weight_set + i);
 		}
 	}
 	else
@@ -1084,8 +1083,7 @@ void compute_quantized_weights_for_decimation(
 
 			// Invert the weight-scaling that was done initially
 			storea(ixl * rscalev + low_boundv, weight_set_out + i);
-			vint scn = pack_low_bytes(weight);
-			store_nbytes(scn, quantized_weight_set + i);
+			pack_and_store_low_bytes(weight, quantized_weight_set + i);
 		}
 	}
 }

Source/astcenc_mathlib.h

@@ -73,10 +73,14 @@
   #endif
 #endif
 
+#ifndef ASTCENC_SVE
+  #define ASTCENC_SVE 0
+#endif
+
 // Force vector-sized SIMD alignment
-#if ASTCENC_AVX
+#if ASTCENC_AVX || ASTCENC_SVE == 8
   #define ASTCENC_VECALIGN 32
-#elif ASTCENC_SSE || ASTCENC_NEON
+#elif ASTCENC_SSE || ASTCENC_NEON || ASTCENC_SVE == 4
   #define ASTCENC_VECALIGN 16
 // Use default alignment for non-SIMD builds
 #else