From 02e3467b75a6e205c6bd2e82d3a1c8daa3c81f6f Mon Sep 17 00:00:00 2001
From: Anthony Blake <anthonix@me.com>
Date: Thu, 15 Nov 2012 16:04:56 +1300
Subject: 2D real-valued transforms work

---
 src/ffts_nd.c      |  14 +++---
 src/ffts_real_nd.c | 130 +++++++----------------------------------------------
 2 files changed, 23 insertions(+), 121 deletions(-)

(limited to 'src')

diff --git a/src/ffts_nd.c b/src/ffts_nd.c
index b87822e..76d3fc4 100644
--- a/src/ffts_nd.c
+++ b/src/ffts_nd.c
@@ -133,13 +133,13 @@ void ffts_transpose(uint64_t *in, uint64_t *out, int w, int h, uint64_t *buf) {
 	size_t i,j;
 	for(i=0;i<w;i+=2) {
 		for(j=0;j<h;j+=2) {
-//			out[i*h + j] = in[j*w + i];
-			__m128d q0 = _mm_load_pd((double *)(in + j*w + i));
-			__m128d q1 = _mm_load_pd((double *)(in + j*w + i + w));
-			__m128d t0 = _mm_shuffle_pd(q0, q1, _MM_SHUFFLE2(0, 0));
-			__m128d t1 = _mm_shuffle_pd(q0, q1, _MM_SHUFFLE2(1, 1));
-			_mm_store_pd((double *)(out + i*h + j), t0);
-			_mm_store_pd((double *)(out + i*h + j + h), t1);
+//		out[i*h + j] = in[j*w + i];
+  		__m128d q0 = _mm_load_pd((double *)(in + j*w + i));
+  		__m128d q1 = _mm_load_pd((double *)(in + j*w + i + w));
+  		__m128d t0 = _mm_shuffle_pd(q0, q1, _MM_SHUFFLE2(0, 0));
+  		__m128d t1 = _mm_shuffle_pd(q0, q1, _MM_SHUFFLE2(1, 1));
+  		_mm_store_pd((double *)(out + i*h + j), t0);
+  		_mm_store_pd((double *)(out + i*h + j + h), t1);
 		}
 	}
 #endif
diff --git a/src/ffts_real_nd.c b/src/ffts_real_nd.c
index 3e87af6..ff97562 100644
--- a/src/ffts_real_nd.c
+++ b/src/ffts_real_nd.c
@@ -36,119 +36,21 @@
 #ifdef __ARM_NEON__
 #include "neon.h"
 #endif
-/*
-void ffts_free_nd_real(ffts_plan_t *p) {
-	free(p->Ns);
-	free(p->Ms);
-
-	int i;
-	for(i=0;i<p->rank;i++) {
-		
-		ffts_plan_t *x = p->plans[i];
-		int k;
-		for(k=0;k<i;k++) {
-			if(x == p->plans[k]) x = NULL;
-		}
-		
-		ffts_free(x);
-	}
-
-	free(p->plans);
-	free(p->buf);
-	free(p->transpose_buf);
-	free(p);
-}
 
-void ffts_transpose(uint64_t *in, uint64_t *out, int w, int h, uint64_t *buf) {
+void ffts_scalar_transpose(uint64_t *in, uint64_t *out, int w, int h, uint64_t *buf) {
 
-#ifdef __ARM_NEON__
-	size_t i,j,k;
-	int linebytes = w*8;
-
-	for(j=0;j<h;j+=8) {
-		for(i=0;i<w;i+=8) {
-			neon_transpose_to_buf(in + j*w + i, buf, w);
-
-			uint64_t __attribute__((aligned(32))) *p = out + i*h + j;
-			uint64_t __attribute__((aligned(32))) *pbuf = buf;
-			uint64_t *ptemp;
-
-			__asm__ __volatile__(
-							 "mov %[ptemp], %[p]\n\t"
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t"
-							 "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t" 
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t"
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t"
-							 "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t" 
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t"
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t"
-							 "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t" 
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t"
-							 "add %[p], %[p], %[w], lsl #3\n\t"
-							 "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t"
-							 "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t"
-							 "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t"
-							 "mov %[ptemp], %[p]\n\t" 
-							 "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t"
-							 "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t"
-							 
-						: [p] "+r" (p), [pbuf] "+r" (pbuf), [ptemp] "+r" (ptemp)
-						: [w] "r" (w)
-						: "memory", "q8", "q9", "q10", "q11"
-						);
-//			out[i*h + j] = in[j*w + i];
-		}
-	}
-#else
 	size_t i,j;
-	for(i=0;i<w;i+=2) {
-		for(j=0;j<h;j+=2) {
-//			out[i*h + j] = in[j*w + i];
-			__m128d q0 = _mm_load_pd((double *)(in + j*w + i));
-			__m128d q1 = _mm_load_pd((double *)(in + j*w + i + w));
-			__m128d t0 = _mm_shuffle_pd(q0, q1, _MM_SHUFFLE2(0, 0));
-			__m128d t1 = _mm_shuffle_pd(q0, q1, _MM_SHUFFLE2(1, 1));
-			_mm_store_pd((double *)(out + i*h + j), t0);
-			_mm_store_pd((double *)(out + i*h + j + h), t1);
+	for(i=0;i<w;i+=1) {
+		for(j=0;j<h;j+=1) {
+			out[i*h + j] = in[j*w + i];
 		}
 	}
-#endif
 
 }
 
-void ffts_execute_nd_real(ffts_plan_t *p, const data_t *  in, data_t *  out) {
+void ffts_execute_nd_real(ffts_plan_t *p, const void *  in, void *  out) {
 
-	uint64_t *din = (uint64_t *)in;
+	uint32_t *din = (uint32_t *)in;
 	uint64_t *buf = p->buf;
 	uint64_t *dout = (uint64_t *)out;
 
@@ -156,22 +58,22 @@ void ffts_execute_nd_real(ffts_plan_t *p, const data_t *  in, data_t *  out) {
 	for(i=0;i<p->Ns[0];i++) {
 		p->plans[0]->transform(p->plans[0], din + (i * p->Ms[0]), buf + (i * p->Ms[0]));	
 	}
-	ffts_transpose(buf, dout, p->Ms[0], p->Ns[0], p->transpose_buf);	
+	ffts_scalar_transpose(buf, dout, p->Ms[0], p->Ns[0], p->transpose_buf);	
 
 	for(i=1;i<p->rank;i++) {
 		for(j=0;j<p->Ns[i];j++) { 
 			p->plans[i]->transform(p->plans[i], dout + (j * p->Ms[i]), buf + (j * p->Ms[i]));	
 		}
-		ffts_transpose(buf, dout, p->Ms[i], p->Ns[i], p->transpose_buf);	
+		ffts_scalar_transpose(buf, dout, p->Ms[i], p->Ns[i], p->transpose_buf);	
 	}
 }
-*/
+
 ffts_plan_t *ffts_init_nd_real(int rank, size_t *Ns, int sign) {
 	size_t vol = 1;
 
 	ffts_plan_t *p = malloc(sizeof(ffts_plan_t));
 
-	p->transform = &ffts_execute_nd;
+	p->transform = &ffts_execute_nd_real;
 	p->destroy = &ffts_free_nd;
 
 	p->rank = rank;
@@ -190,13 +92,13 @@ ffts_plan_t *ffts_init_nd_real(int rank, size_t *Ns, int sign) {
 
 		p->plans[i] = NULL;
 		int k;
-		for(k=1;k<i;k++) {
-			if(p->Ms[k] == p->Ms[i]) 
-				p->plans[i] = p->plans[k];
-		}
+  	for(k=1;k<i;k++) {
+  		if(p->Ms[k] == p->Ms[i]) 
+  			p->plans[i] = p->plans[k];
+  	}
 
-		if(!i)           p->plans[i] = ffts_init_1d_real(p->Ms[i], sign); 
-		if(!p->plans[i]) p->plans[i] = ffts_init_1d(p->Ms[i], sign); 
+		if(!i)                p->plans[i] = ffts_init_1d_real(p->Ms[i], sign); 
+		else if(!p->plans[i]) p->plans[i] = ffts_init_1d(p->Ms[i], sign); 
 	
 	}
 
-- 
cgit v1.1