Vulkan: Add DP4A MMQ and Q8_1 quantization shader

Author: 0cc4m

ggml/src/ggml-quants.c

@@ -2020,6 +2020,13 @@ size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nr
     return nrow * row_size;
 }
 
+size_t quantize_q8_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    (void)quant_weights; // not used
+    const size_t row_size = ggml_row_size(GGML_TYPE_Q8_1, n_per_row);
+    quantize_row_q8_1_ref(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * row_size;
+}
+
 // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
 
 void quantize_row_tq1_0_ref(const float * restrict x, block_tq1_0 * restrict y, int64_t k) {

ggml/src/ggml-quants.h

@@ -89,6 +89,7 @@ GGML_API size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q8_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 GGML_API void iq2xs_init_impl(enum ggml_type type);
 GGML_API void iq2xs_free_impl(enum ggml_type type);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -212,7 +212,7 @@ void main() {
 #else
     ACC_TYPE sums[WMITER * TM * WNITER * TN];
     FLOAT_TYPE cache_a[WMITER * TM];
-    FLOAT_TYPE cache_b[WNITER * TN];
+    FLOAT_TYPE cache_b[TN];
 
     [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
         sums[i] = ACC_TYPE(0.0f);
@@ -744,16 +744,14 @@ void main() {
             }
             [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                 [[unroll]] for (uint j = 0; j < TN; j++) {
-                    cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
+                    cache_b[j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
                 }
-            }
 
-            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                 [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                     [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                         [[unroll]] for (uint cr = 0; cr < TM; cr++) {
                             const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[wsic * TN + cc]), sums[sums_idx]);
+                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[cc]), sums[sums_idx]);
                         }
                     }
                 }