Attempt matmul vec but failing

Author: cirosantilli

gcc/README.md

@@ -26,6 +26,7 @@ Cheat on the GNU Compile Collection (GCC) language extensions and command line u
     1.  [Attribute](attribute.c)
         1.  Variable
             1.  [weak](weak/)
+            1.  [Vector extensions](vector.c)
         1.  Function
             1.  [sentinel](sentinel.c)
     1.  [asm](asm.c)

gcc/vector.c

@@ -0,0 +1,64 @@
+/*
+Increase the chance that the compiler will use SIMD instructions.
+
+Basically a more abstract and ISA portable (TODO?) version of intrinsics.
+
+Note however that the GCC and other compiles have an auto-vectorization
+optimization, which might use SIMD even if you don't use the vector extensions.
+*/
+
+#include "common.h"
+
+typedef int int4 __attribute__ ((vector_size(4 * sizeof(int))));
+
+/*
+TODO: SIMD to scalar operations don't appear possible without intrinsics:
+http://stackoverflow.com/questions/31597302/gcc-c-vector-extension-how-to-check-if-result-of-any-element-wise-comparison-is
+*/
+int int4_eq(int4 i, int4 j) {
+	return
+		i[0] == j[0] &&
+		i[1] == j[1] &&
+		i[2] == j[2] &&
+		i[3] == j[3]
+	;
+}
+
+int main(void) {
+	/* Hello world. */
+	{
+		int4 i = {1, 2, 3, 4};
+		int4 j = {5, 6, 7, 8};
+		int4 k = i + j;
+		assert(int4_eq(k, (int4){6, 8, 10, 12}));
+	}
+
+	/* From / to array. TODO any better way than going over all indices?
+	 * Intrinsic way: _mm_load_pd family.
+	 * */
+	{
+		int is[] = {1, 2, 3, 4};
+		int js[] = {5, 6, 7, 8};
+		int ks[4];
+		int ks2[] = {6, 8, 10, 12};
+		int4 i;
+		int4 j;
+		int4 k;
+		i[0] = is[0];
+		i[1] = is[1];
+		i[2] = is[2];
+		i[3] = is[3];
+		j[0] = js[0];
+		j[1] = js[1];
+		j[2] = js[2];
+		j[3] = js[3];
+		k = i + j;
+		ks[0] = k[0];
+		ks[1] = k[1];
+		ks[2] = k[2];
+		ks[3] = k[3];
+		assert(memcmp(ks, ks2, sizeof(ks)) == 0);
+	}
+
+    return EXIT_SUCCESS;
+}

opencl/README.md

@@ -9,7 +9,9 @@
     1.  [Work item built-ins](work_item_builtin.c)
     1.  [Vector type](vector_type.c)
     1.  [clinfo](clinfo.c)
-    1.  [Matrix multiplication](matmul.c)
+    1.  [Preprocessor](matmul.c)
+    1.  [Benchmark examples](benchmark-examples.md)
+        1.  [Matrix multiplication](matmul.c)
 1.  Tools
     1.  [clinfo](clinfo.md)
     1.  [Benchmarks](benchmarks.md)

opencl/benchmark-examples.md

@@ -0,0 +1,5 @@
+# Benchmark examples
+
+Examples under this section do benchmarks, and thus take longer to finish.
+
+This is the case of all useful examples, i.e. that actually run a real-ish application faster per dollar than the best CPU implementation we can craft, including intrinsics.

opencl/common.h

@@ -23,9 +23,10 @@ typedef struct {
     cl_program program;
 } Common;
 
-void common_init(
+void common_init_options(
     Common *common,
-    const char *source
+    const char *source,
+    const char *options
 ) {
 	char *err;
 	size_t err_len;
@@ -36,7 +37,7 @@ void common_init(
     clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 1, &(common->device), NULL);
     common->context = clCreateContext(NULL, 1, &(common->device), NULL, NULL, NULL);
     common->program = clCreateProgramWithSource(common->context, 1, &source, NULL, NULL);
-    ret = clBuildProgram(common->program, 1, &(common->device), "", NULL, NULL);
+    ret = clBuildProgram(common->program, 1, &(common->device), options, NULL, NULL);
     if (CL_SUCCESS != ret) {
 		clGetProgramBuildInfo(common->program, common->device, CL_PROGRAM_BUILD_LOG, 0, NULL, &err_len);
 		err = malloc(err_len);
@@ -49,6 +50,13 @@ void common_init(
     common->command_queue = clCreateCommandQueue(common->context, common->device, 0, NULL);
 }
 
+void common_init(
+    Common *common,
+    const char *source
+) {
+	common_init_options(common, source, "");
+}
+
 char* common_read_file(const char *path) {
     char *buffer;
     FILE *f;

opencl/inc.c

@@ -9,6 +9,7 @@ This is our OpenCL hello world, so we are not doing:
 
 #include <assert.h>
 #include <stdio.h>
+#include <stdlib.h>
 
 /* To prevent deprecation warnings when headers are 2.0. */
 #define CL_USE_DEPRECATED_OPENCL_1_2_APIS